1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where algebraic
12 // simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Target/TargetData.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Support/CallSite.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/GetElementPtrTypeIterator.h"
48 #include "llvm/Support/InstVisitor.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/PatternMatch.h"
51 #include "llvm/Support/Compiler.h"
52 #include "llvm/ADT/Statistic.h"
53 #include "llvm/ADT/STLExtras.h"
56 using namespace llvm::PatternMatch;
58 STATISTIC(NumCombined , "Number of insts combined");
59 STATISTIC(NumConstProp, "Number of constant folds");
60 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
61 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
62 STATISTIC(NumSunkInst , "Number of instructions sunk");
65 class VISIBILITY_HIDDEN InstCombiner
66 : public FunctionPass,
67 public InstVisitor<InstCombiner, Instruction*> {
68 // Worklist of all of the instructions that need to be simplified.
69 std::vector<Instruction*> WorkList;
72 /// AddUsersToWorkList - When an instruction is simplified, add all users of
73 /// the instruction to the work lists because they might get more simplified
76 void AddUsersToWorkList(Value &I) {
77 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
79 WorkList.push_back(cast<Instruction>(*UI));
82 /// AddUsesToWorkList - When an instruction is simplified, add operands to
83 /// the work lists because they might get more simplified now.
85 void AddUsesToWorkList(Instruction &I) {
86 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
87 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
88 WorkList.push_back(Op);
91 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
92 /// dead. Add all of its operands to the worklist, turning them into
93 /// undef's to reduce the number of uses of those instructions.
95 /// Return the specified operand before it is turned into an undef.
97 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
98 Value *R = I.getOperand(op);
100 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
101 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
102 WorkList.push_back(Op);
103 // Set the operand to undef to drop the use.
104 I.setOperand(i, UndefValue::get(Op->getType()));
110 // removeFromWorkList - remove all instances of I from the worklist.
111 void removeFromWorkList(Instruction *I);
113 virtual bool runOnFunction(Function &F);
115 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
116 AU.addRequired<TargetData>();
117 AU.addPreservedID(LCSSAID);
118 AU.setPreservesCFG();
121 TargetData &getTargetData() const { return *TD; }
123 // Visitation implementation - Implement instruction combining for different
124 // instruction types. The semantics are as follows:
126 // null - No change was made
127 // I - Change was made, I is still valid, I may be dead though
128 // otherwise - Change was made, replace I with returned instruction
130 Instruction *visitAdd(BinaryOperator &I);
131 Instruction *visitSub(BinaryOperator &I);
132 Instruction *visitMul(BinaryOperator &I);
133 Instruction *visitURem(BinaryOperator &I);
134 Instruction *visitSRem(BinaryOperator &I);
135 Instruction *visitFRem(BinaryOperator &I);
136 Instruction *commonRemTransforms(BinaryOperator &I);
137 Instruction *commonIRemTransforms(BinaryOperator &I);
138 Instruction *commonDivTransforms(BinaryOperator &I);
139 Instruction *commonIDivTransforms(BinaryOperator &I);
140 Instruction *visitUDiv(BinaryOperator &I);
141 Instruction *visitSDiv(BinaryOperator &I);
142 Instruction *visitFDiv(BinaryOperator &I);
143 Instruction *visitAnd(BinaryOperator &I);
144 Instruction *visitOr (BinaryOperator &I);
145 Instruction *visitXor(BinaryOperator &I);
146 Instruction *visitFCmpInst(FCmpInst &I);
147 Instruction *visitICmpInst(ICmpInst &I);
148 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
150 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
151 ICmpInst::Predicate Cond, Instruction &I);
152 Instruction *visitShiftInst(ShiftInst &I);
153 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
155 Instruction *commonCastTransforms(CastInst &CI);
156 Instruction *commonIntCastTransforms(CastInst &CI);
157 Instruction *visitTrunc(CastInst &CI);
158 Instruction *visitZExt(CastInst &CI);
159 Instruction *visitSExt(CastInst &CI);
160 Instruction *visitFPTrunc(CastInst &CI);
161 Instruction *visitFPExt(CastInst &CI);
162 Instruction *visitFPToUI(CastInst &CI);
163 Instruction *visitFPToSI(CastInst &CI);
164 Instruction *visitUIToFP(CastInst &CI);
165 Instruction *visitSIToFP(CastInst &CI);
166 Instruction *visitPtrToInt(CastInst &CI);
167 Instruction *visitIntToPtr(CastInst &CI);
168 Instruction *visitBitCast(CastInst &CI);
169 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
171 Instruction *visitSelectInst(SelectInst &CI);
172 Instruction *visitCallInst(CallInst &CI);
173 Instruction *visitInvokeInst(InvokeInst &II);
174 Instruction *visitPHINode(PHINode &PN);
175 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
176 Instruction *visitAllocationInst(AllocationInst &AI);
177 Instruction *visitFreeInst(FreeInst &FI);
178 Instruction *visitLoadInst(LoadInst &LI);
179 Instruction *visitStoreInst(StoreInst &SI);
180 Instruction *visitBranchInst(BranchInst &BI);
181 Instruction *visitSwitchInst(SwitchInst &SI);
182 Instruction *visitInsertElementInst(InsertElementInst &IE);
183 Instruction *visitExtractElementInst(ExtractElementInst &EI);
184 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
186 // visitInstruction - Specify what to return for unhandled instructions...
187 Instruction *visitInstruction(Instruction &I) { return 0; }
190 Instruction *visitCallSite(CallSite CS);
191 bool transformConstExprCastCall(CallSite CS);
194 // InsertNewInstBefore - insert an instruction New before instruction Old
195 // in the program. Add the new instruction to the worklist.
197 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
198 assert(New && New->getParent() == 0 &&
199 "New instruction already inserted into a basic block!");
200 BasicBlock *BB = Old.getParent();
201 BB->getInstList().insert(&Old, New); // Insert inst
202 WorkList.push_back(New); // Add to worklist
206 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
207 /// This also adds the cast to the worklist. Finally, this returns the
209 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
211 if (V->getType() == Ty) return V;
213 if (Constant *CV = dyn_cast<Constant>(V))
214 return ConstantExpr::getCast(opc, CV, Ty);
216 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
217 WorkList.push_back(C);
221 // ReplaceInstUsesWith - This method is to be used when an instruction is
222 // found to be dead, replacable with another preexisting expression. Here
223 // we add all uses of I to the worklist, replace all uses of I with the new
224 // value, then return I, so that the inst combiner will know that I was
227 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
228 AddUsersToWorkList(I); // Add all modified instrs to worklist
230 I.replaceAllUsesWith(V);
233 // If we are replacing the instruction with itself, this must be in a
234 // segment of unreachable code, so just clobber the instruction.
235 I.replaceAllUsesWith(UndefValue::get(I.getType()));
240 // UpdateValueUsesWith - This method is to be used when an value is
241 // found to be replacable with another preexisting expression or was
242 // updated. Here we add all uses of I to the worklist, replace all uses of
243 // I with the new value (unless the instruction was just updated), then
244 // return true, so that the inst combiner will know that I was modified.
246 bool UpdateValueUsesWith(Value *Old, Value *New) {
247 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
249 Old->replaceAllUsesWith(New);
250 if (Instruction *I = dyn_cast<Instruction>(Old))
251 WorkList.push_back(I);
252 if (Instruction *I = dyn_cast<Instruction>(New))
253 WorkList.push_back(I);
257 // EraseInstFromFunction - When dealing with an instruction that has side
258 // effects or produces a void value, we can't rely on DCE to delete the
259 // instruction. Instead, visit methods should return the value returned by
261 Instruction *EraseInstFromFunction(Instruction &I) {
262 assert(I.use_empty() && "Cannot erase instruction that is used!");
263 AddUsesToWorkList(I);
264 removeFromWorkList(&I);
266 return 0; // Don't do anything with FI
270 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
271 /// InsertBefore instruction. This is specialized a bit to avoid inserting
272 /// casts that are known to not do anything...
274 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
275 Value *V, const Type *DestTy,
276 Instruction *InsertBefore);
278 /// SimplifyCommutative - This performs a few simplifications for
279 /// commutative operators.
280 bool SimplifyCommutative(BinaryOperator &I);
282 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
283 /// most-complex to least-complex order.
284 bool SimplifyCompare(CmpInst &I);
286 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
287 uint64_t &KnownZero, uint64_t &KnownOne,
290 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
291 uint64_t &UndefElts, unsigned Depth = 0);
293 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
294 // PHI node as operand #0, see if we can fold the instruction into the PHI
295 // (which is only possible if all operands to the PHI are constants).
296 Instruction *FoldOpIntoPhi(Instruction &I);
298 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
299 // operator and they all are only used by the PHI, PHI together their
300 // inputs, and do the operation once, to the result of the PHI.
301 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
302 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
305 Instruction *OptAndOp(Instruction *Op, ConstantIntegral *OpRHS,
306 ConstantIntegral *AndRHS, BinaryOperator &TheAnd);
308 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantIntegral *Mask,
309 bool isSub, Instruction &I);
310 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
311 bool isSigned, bool Inside, Instruction &IB);
312 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
313 Instruction *MatchBSwap(BinaryOperator &I);
315 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
318 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
321 // getComplexity: Assign a complexity or rank value to LLVM Values...
322 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
323 static unsigned getComplexity(Value *V) {
324 if (isa<Instruction>(V)) {
325 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
329 if (isa<Argument>(V)) return 3;
330 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
333 // isOnlyUse - Return true if this instruction will be deleted if we stop using
335 static bool isOnlyUse(Value *V) {
336 return V->hasOneUse() || isa<Constant>(V);
339 // getPromotedType - Return the specified type promoted as it would be to pass
340 // though a va_arg area...
341 static const Type *getPromotedType(const Type *Ty) {
342 switch (Ty->getTypeID()) {
343 case Type::SByteTyID:
344 case Type::ShortTyID: return Type::IntTy;
345 case Type::UByteTyID:
346 case Type::UShortTyID: return Type::UIntTy;
347 case Type::FloatTyID: return Type::DoubleTy;
352 /// getBitCastOperand - If the specified operand is a CastInst or a constant
353 /// expression bitcast, return the operand value, otherwise return null.
354 static Value *getBitCastOperand(Value *V) {
355 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
356 return I->getOperand(0);
357 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
358 if (CE->getOpcode() == Instruction::BitCast)
359 return CE->getOperand(0);
363 /// This function is a wrapper around CastInst::isEliminableCastPair. It
364 /// simply extracts arguments and returns what that function returns.
365 /// @Determine if it is valid to eliminate a Convert pair
366 static Instruction::CastOps
367 isEliminableCastPair(
368 const CastInst *CI, ///< The first cast instruction
369 unsigned opcode, ///< The opcode of the second cast instruction
370 const Type *DstTy, ///< The target type for the second cast instruction
371 TargetData *TD ///< The target data for pointer size
374 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
375 const Type *MidTy = CI->getType(); // B from above
377 // Get the opcodes of the two Cast instructions
378 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
379 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
381 return Instruction::CastOps(
382 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
383 DstTy, TD->getIntPtrType()));
386 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
387 /// in any code being generated. It does not require codegen if V is simple
388 /// enough or if the cast can be folded into other casts.
389 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
390 const Type *Ty, TargetData *TD) {
391 if (V->getType() == Ty || isa<Constant>(V)) return false;
393 // If this is a noop cast, it isn't real codegen.
394 if (V->getType()->canLosslesslyBitCastTo(Ty))
397 // If this is another cast that can be eliminated, it isn't codegen either.
398 if (const CastInst *CI = dyn_cast<CastInst>(V))
399 if (isEliminableCastPair(CI, opcode, Ty, TD))
404 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
405 /// InsertBefore instruction. This is specialized a bit to avoid inserting
406 /// casts that are known to not do anything...
408 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
409 Value *V, const Type *DestTy,
410 Instruction *InsertBefore) {
411 if (V->getType() == DestTy) return V;
412 if (Constant *C = dyn_cast<Constant>(V))
413 return ConstantExpr::getCast(opcode, C, DestTy);
415 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
418 // SimplifyCommutative - This performs a few simplifications for commutative
421 // 1. Order operands such that they are listed from right (least complex) to
422 // left (most complex). This puts constants before unary operators before
425 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
426 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
428 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
429 bool Changed = false;
430 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
431 Changed = !I.swapOperands();
433 if (!I.isAssociative()) return Changed;
434 Instruction::BinaryOps Opcode = I.getOpcode();
435 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
436 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
437 if (isa<Constant>(I.getOperand(1))) {
438 Constant *Folded = ConstantExpr::get(I.getOpcode(),
439 cast<Constant>(I.getOperand(1)),
440 cast<Constant>(Op->getOperand(1)));
441 I.setOperand(0, Op->getOperand(0));
442 I.setOperand(1, Folded);
444 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
445 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
446 isOnlyUse(Op) && isOnlyUse(Op1)) {
447 Constant *C1 = cast<Constant>(Op->getOperand(1));
448 Constant *C2 = cast<Constant>(Op1->getOperand(1));
450 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
451 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
452 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
455 WorkList.push_back(New);
456 I.setOperand(0, New);
457 I.setOperand(1, Folded);
464 /// SimplifyCompare - For a CmpInst this function just orders the operands
465 /// so that theyare listed from right (least complex) to left (most complex).
466 /// This puts constants before unary operators before binary operators.
467 bool InstCombiner::SimplifyCompare(CmpInst &I) {
468 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
471 // Compare instructions are not associative so there's nothing else we can do.
475 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
476 // if the LHS is a constant zero (which is the 'negate' form).
478 static inline Value *dyn_castNegVal(Value *V) {
479 if (BinaryOperator::isNeg(V))
480 return BinaryOperator::getNegArgument(V);
482 // Constants can be considered to be negated values if they can be folded.
483 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
484 return ConstantExpr::getNeg(C);
488 static inline Value *dyn_castNotVal(Value *V) {
489 if (BinaryOperator::isNot(V))
490 return BinaryOperator::getNotArgument(V);
492 // Constants can be considered to be not'ed values...
493 if (ConstantIntegral *C = dyn_cast<ConstantIntegral>(V))
494 return ConstantExpr::getNot(C);
498 // dyn_castFoldableMul - If this value is a multiply that can be folded into
499 // other computations (because it has a constant operand), return the
500 // non-constant operand of the multiply, and set CST to point to the multiplier.
501 // Otherwise, return null.
503 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
504 if (V->hasOneUse() && V->getType()->isInteger())
505 if (Instruction *I = dyn_cast<Instruction>(V)) {
506 if (I->getOpcode() == Instruction::Mul)
507 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
508 return I->getOperand(0);
509 if (I->getOpcode() == Instruction::Shl)
510 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
511 // The multiplier is really 1 << CST.
512 Constant *One = ConstantInt::get(V->getType(), 1);
513 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
514 return I->getOperand(0);
520 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
521 /// expression, return it.
522 static User *dyn_castGetElementPtr(Value *V) {
523 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
524 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
525 if (CE->getOpcode() == Instruction::GetElementPtr)
526 return cast<User>(V);
530 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
531 static ConstantInt *AddOne(ConstantInt *C) {
532 return cast<ConstantInt>(ConstantExpr::getAdd(C,
533 ConstantInt::get(C->getType(), 1)));
535 static ConstantInt *SubOne(ConstantInt *C) {
536 return cast<ConstantInt>(ConstantExpr::getSub(C,
537 ConstantInt::get(C->getType(), 1)));
540 /// GetConstantInType - Return a ConstantInt with the specified type and value.
542 static ConstantIntegral *GetConstantInType(const Type *Ty, uint64_t Val) {
543 if (Ty->isUnsigned())
544 return ConstantInt::get(Ty, Val);
545 else if (Ty->getTypeID() == Type::BoolTyID)
546 return ConstantBool::get(Val);
548 SVal <<= 64-Ty->getPrimitiveSizeInBits();
549 SVal >>= 64-Ty->getPrimitiveSizeInBits();
550 return ConstantInt::get(Ty, SVal);
554 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
555 /// known to be either zero or one and return them in the KnownZero/KnownOne
556 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
558 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
559 uint64_t &KnownOne, unsigned Depth = 0) {
560 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
561 // we cannot optimize based on the assumption that it is zero without changing
562 // it to be an explicit zero. If we don't change it to zero, other code could
563 // optimized based on the contradictory assumption that it is non-zero.
564 // Because instcombine aggressively folds operations with undef args anyway,
565 // this won't lose us code quality.
566 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
567 // We know all of the bits for a constant!
568 KnownOne = CI->getZExtValue() & Mask;
569 KnownZero = ~KnownOne & Mask;
573 KnownZero = KnownOne = 0; // Don't know anything.
574 if (Depth == 6 || Mask == 0)
575 return; // Limit search depth.
577 uint64_t KnownZero2, KnownOne2;
578 Instruction *I = dyn_cast<Instruction>(V);
581 Mask &= V->getType()->getIntegralTypeMask();
583 switch (I->getOpcode()) {
584 case Instruction::And:
585 // If either the LHS or the RHS are Zero, the result is zero.
586 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
588 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
589 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
590 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
592 // Output known-1 bits are only known if set in both the LHS & RHS.
593 KnownOne &= KnownOne2;
594 // Output known-0 are known to be clear if zero in either the LHS | RHS.
595 KnownZero |= KnownZero2;
597 case Instruction::Or:
598 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
600 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
601 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
602 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
604 // Output known-0 bits are only known if clear in both the LHS & RHS.
605 KnownZero &= KnownZero2;
606 // Output known-1 are known to be set if set in either the LHS | RHS.
607 KnownOne |= KnownOne2;
609 case Instruction::Xor: {
610 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
611 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
612 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
613 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
615 // Output known-0 bits are known if clear or set in both the LHS & RHS.
616 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
617 // Output known-1 are known to be set if set in only one of the LHS, RHS.
618 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
619 KnownZero = KnownZeroOut;
622 case Instruction::Select:
623 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
624 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
625 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
626 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
628 // Only known if known in both the LHS and RHS.
629 KnownOne &= KnownOne2;
630 KnownZero &= KnownZero2;
632 case Instruction::FPTrunc:
633 case Instruction::FPExt:
634 case Instruction::FPToUI:
635 case Instruction::FPToSI:
636 case Instruction::SIToFP:
637 case Instruction::PtrToInt:
638 case Instruction::UIToFP:
639 case Instruction::IntToPtr:
640 return; // Can't work with floating point or pointers
641 case Instruction::Trunc:
642 // All these have integer operands
643 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
645 case Instruction::BitCast: {
646 const Type *SrcTy = I->getOperand(0)->getType();
647 if (SrcTy->isIntegral()) {
648 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
653 case Instruction::ZExt: {
654 // Compute the bits in the result that are not present in the input.
655 const Type *SrcTy = I->getOperand(0)->getType();
656 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
657 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
659 Mask &= SrcTy->getIntegralTypeMask();
660 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
661 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
662 // The top bits are known to be zero.
663 KnownZero |= NewBits;
666 case Instruction::SExt: {
667 // Compute the bits in the result that are not present in the input.
668 const Type *SrcTy = I->getOperand(0)->getType();
669 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
670 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
672 Mask &= SrcTy->getIntegralTypeMask();
673 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
674 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
676 // If the sign bit of the input is known set or clear, then we know the
677 // top bits of the result.
678 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
679 if (KnownZero & InSignBit) { // Input sign bit known zero
680 KnownZero |= NewBits;
681 KnownOne &= ~NewBits;
682 } else if (KnownOne & InSignBit) { // Input sign bit known set
684 KnownZero &= ~NewBits;
685 } else { // Input sign bit unknown
686 KnownZero &= ~NewBits;
687 KnownOne &= ~NewBits;
691 case Instruction::Shl:
692 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
693 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
694 uint64_t ShiftAmt = SA->getZExtValue();
696 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
697 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
698 KnownZero <<= ShiftAmt;
699 KnownOne <<= ShiftAmt;
700 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
704 case Instruction::LShr:
705 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
706 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
707 // Compute the new bits that are at the top now.
708 uint64_t ShiftAmt = SA->getZExtValue();
709 uint64_t HighBits = (1ULL << ShiftAmt)-1;
710 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
712 // Unsigned shift right.
714 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
715 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
716 KnownZero >>= ShiftAmt;
717 KnownOne >>= ShiftAmt;
718 KnownZero |= HighBits; // high bits known zero.
722 case Instruction::AShr:
723 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
724 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
725 // Compute the new bits that are at the top now.
726 uint64_t ShiftAmt = SA->getZExtValue();
727 uint64_t HighBits = (1ULL << ShiftAmt)-1;
728 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
730 // Signed shift right.
732 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
733 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
734 KnownZero >>= ShiftAmt;
735 KnownOne >>= ShiftAmt;
737 // Handle the sign bits.
738 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
739 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
741 if (KnownZero & SignBit) { // New bits are known zero.
742 KnownZero |= HighBits;
743 } else if (KnownOne & SignBit) { // New bits are known one.
744 KnownOne |= HighBits;
752 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
753 /// this predicate to simplify operations downstream. Mask is known to be zero
754 /// for bits that V cannot have.
755 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
756 uint64_t KnownZero, KnownOne;
757 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
758 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
759 return (KnownZero & Mask) == Mask;
762 /// ShrinkDemandedConstant - Check to see if the specified operand of the
763 /// specified instruction is a constant integer. If so, check to see if there
764 /// are any bits set in the constant that are not demanded. If so, shrink the
765 /// constant and return true.
766 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
768 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
769 if (!OpC) return false;
771 // If there are no bits set that aren't demanded, nothing to do.
772 if ((~Demanded & OpC->getZExtValue()) == 0)
775 // This is producing any bits that are not needed, shrink the RHS.
776 uint64_t Val = Demanded & OpC->getZExtValue();
777 I->setOperand(OpNo, GetConstantInType(OpC->getType(), Val));
781 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
782 // set of known zero and one bits, compute the maximum and minimum values that
783 // could have the specified known zero and known one bits, returning them in
785 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
788 int64_t &Min, int64_t &Max) {
789 uint64_t TypeBits = Ty->getIntegralTypeMask();
790 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
792 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
794 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
795 // bit if it is unknown.
797 Max = KnownOne|UnknownBits;
799 if (SignBit & UnknownBits) { // Sign bit is unknown
804 // Sign extend the min/max values.
805 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
806 Min = (Min << ShAmt) >> ShAmt;
807 Max = (Max << ShAmt) >> ShAmt;
810 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
811 // a set of known zero and one bits, compute the maximum and minimum values that
812 // could have the specified known zero and known one bits, returning them in
814 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
819 uint64_t TypeBits = Ty->getIntegralTypeMask();
820 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
822 // The minimum value is when the unknown bits are all zeros.
824 // The maximum value is when the unknown bits are all ones.
825 Max = KnownOne|UnknownBits;
829 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
830 /// DemandedMask bits of the result of V are ever used downstream. If we can
831 /// use this information to simplify V, do so and return true. Otherwise,
832 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
833 /// the expression (used to simplify the caller). The KnownZero/One bits may
834 /// only be accurate for those bits in the DemandedMask.
835 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
836 uint64_t &KnownZero, uint64_t &KnownOne,
838 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
839 // We know all of the bits for a constant!
840 KnownOne = CI->getZExtValue() & DemandedMask;
841 KnownZero = ~KnownOne & DemandedMask;
845 KnownZero = KnownOne = 0;
846 if (!V->hasOneUse()) { // Other users may use these bits.
847 if (Depth != 0) { // Not at the root.
848 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
849 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
852 // If this is the root being simplified, allow it to have multiple uses,
853 // just set the DemandedMask to all bits.
854 DemandedMask = V->getType()->getIntegralTypeMask();
855 } else if (DemandedMask == 0) { // Not demanding any bits from V.
856 if (V != UndefValue::get(V->getType()))
857 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
859 } else if (Depth == 6) { // Limit search depth.
863 Instruction *I = dyn_cast<Instruction>(V);
864 if (!I) return false; // Only analyze instructions.
866 DemandedMask &= V->getType()->getIntegralTypeMask();
868 uint64_t KnownZero2 = 0, KnownOne2 = 0;
869 switch (I->getOpcode()) {
871 case Instruction::And:
872 // If either the LHS or the RHS are Zero, the result is zero.
873 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
874 KnownZero, KnownOne, Depth+1))
876 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
878 // If something is known zero on the RHS, the bits aren't demanded on the
880 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
881 KnownZero2, KnownOne2, Depth+1))
883 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
885 // If all of the demanded bits are known 1 on one side, return the other.
886 // These bits cannot contribute to the result of the 'and'.
887 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
888 return UpdateValueUsesWith(I, I->getOperand(0));
889 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
890 return UpdateValueUsesWith(I, I->getOperand(1));
892 // If all of the demanded bits in the inputs are known zeros, return zero.
893 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
894 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
896 // If the RHS is a constant, see if we can simplify it.
897 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
898 return UpdateValueUsesWith(I, I);
900 // Output known-1 bits are only known if set in both the LHS & RHS.
901 KnownOne &= KnownOne2;
902 // Output known-0 are known to be clear if zero in either the LHS | RHS.
903 KnownZero |= KnownZero2;
905 case Instruction::Or:
906 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
907 KnownZero, KnownOne, Depth+1))
909 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
910 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
911 KnownZero2, KnownOne2, Depth+1))
913 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
915 // If all of the demanded bits are known zero on one side, return the other.
916 // These bits cannot contribute to the result of the 'or'.
917 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
918 return UpdateValueUsesWith(I, I->getOperand(0));
919 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
920 return UpdateValueUsesWith(I, I->getOperand(1));
922 // If all of the potentially set bits on one side are known to be set on
923 // the other side, just use the 'other' side.
924 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
925 (DemandedMask & (~KnownZero)))
926 return UpdateValueUsesWith(I, I->getOperand(0));
927 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
928 (DemandedMask & (~KnownZero2)))
929 return UpdateValueUsesWith(I, I->getOperand(1));
931 // If the RHS is a constant, see if we can simplify it.
932 if (ShrinkDemandedConstant(I, 1, DemandedMask))
933 return UpdateValueUsesWith(I, I);
935 // Output known-0 bits are only known if clear in both the LHS & RHS.
936 KnownZero &= KnownZero2;
937 // Output known-1 are known to be set if set in either the LHS | RHS.
938 KnownOne |= KnownOne2;
940 case Instruction::Xor: {
941 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
942 KnownZero, KnownOne, Depth+1))
944 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
945 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
946 KnownZero2, KnownOne2, Depth+1))
948 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
950 // If all of the demanded bits are known zero on one side, return the other.
951 // These bits cannot contribute to the result of the 'xor'.
952 if ((DemandedMask & KnownZero) == DemandedMask)
953 return UpdateValueUsesWith(I, I->getOperand(0));
954 if ((DemandedMask & KnownZero2) == DemandedMask)
955 return UpdateValueUsesWith(I, I->getOperand(1));
957 // Output known-0 bits are known if clear or set in both the LHS & RHS.
958 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
959 // Output known-1 are known to be set if set in only one of the LHS, RHS.
960 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
962 // If all of the demanded bits are known to be zero on one side or the
963 // other, turn this into an *inclusive* or.
964 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
965 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
967 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
969 InsertNewInstBefore(Or, *I);
970 return UpdateValueUsesWith(I, Or);
973 // If all of the demanded bits on one side are known, and all of the set
974 // bits on that side are also known to be set on the other side, turn this
975 // into an AND, as we know the bits will be cleared.
976 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
977 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
978 if ((KnownOne & KnownOne2) == KnownOne) {
979 Constant *AndC = GetConstantInType(I->getType(),
980 ~KnownOne & DemandedMask);
982 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
983 InsertNewInstBefore(And, *I);
984 return UpdateValueUsesWith(I, And);
988 // If the RHS is a constant, see if we can simplify it.
989 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
990 if (ShrinkDemandedConstant(I, 1, DemandedMask))
991 return UpdateValueUsesWith(I, I);
993 KnownZero = KnownZeroOut;
994 KnownOne = KnownOneOut;
997 case Instruction::Select:
998 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
999 KnownZero, KnownOne, Depth+1))
1001 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1002 KnownZero2, KnownOne2, Depth+1))
1004 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1005 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
1007 // If the operands are constants, see if we can simplify them.
1008 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1009 return UpdateValueUsesWith(I, I);
1010 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1011 return UpdateValueUsesWith(I, I);
1013 // Only known if known in both the LHS and RHS.
1014 KnownOne &= KnownOne2;
1015 KnownZero &= KnownZero2;
1017 case Instruction::Trunc:
1018 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1019 KnownZero, KnownOne, Depth+1))
1021 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1023 case Instruction::BitCast:
1024 if (!I->getOperand(0)->getType()->isIntegral())
1027 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1028 KnownZero, KnownOne, Depth+1))
1030 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1032 case Instruction::ZExt: {
1033 // Compute the bits in the result that are not present in the input.
1034 const Type *SrcTy = I->getOperand(0)->getType();
1035 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1036 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1038 DemandedMask &= SrcTy->getIntegralTypeMask();
1039 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1040 KnownZero, KnownOne, Depth+1))
1042 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1043 // The top bits are known to be zero.
1044 KnownZero |= NewBits;
1047 case Instruction::SExt: {
1048 // Compute the bits in the result that are not present in the input.
1049 const Type *SrcTy = I->getOperand(0)->getType();
1050 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1051 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1053 // Get the sign bit for the source type
1054 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1055 int64_t InputDemandedBits = DemandedMask & SrcTy->getIntegralTypeMask();
1057 // If any of the sign extended bits are demanded, we know that the sign
1059 if (NewBits & DemandedMask)
1060 InputDemandedBits |= InSignBit;
1062 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1063 KnownZero, KnownOne, Depth+1))
1065 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1067 // If the sign bit of the input is known set or clear, then we know the
1068 // top bits of the result.
1070 // If the input sign bit is known zero, or if the NewBits are not demanded
1071 // convert this into a zero extension.
1072 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1073 // Convert to ZExt cast
1074 CastInst *NewCast = CastInst::create(
1075 Instruction::ZExt, I->getOperand(0), I->getType(), I->getName(), I);
1076 return UpdateValueUsesWith(I, NewCast);
1077 } else if (KnownOne & InSignBit) { // Input sign bit known set
1078 KnownOne |= NewBits;
1079 KnownZero &= ~NewBits;
1080 } else { // Input sign bit unknown
1081 KnownZero &= ~NewBits;
1082 KnownOne &= ~NewBits;
1086 case Instruction::Add:
1087 // If there is a constant on the RHS, there are a variety of xformations
1089 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1090 // If null, this should be simplified elsewhere. Some of the xforms here
1091 // won't work if the RHS is zero.
1092 if (RHS->isNullValue())
1095 // Figure out what the input bits are. If the top bits of the and result
1096 // are not demanded, then the add doesn't demand them from its input
1099 // Shift the demanded mask up so that it's at the top of the uint64_t.
1100 unsigned BitWidth = I->getType()->getPrimitiveSizeInBits();
1101 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1103 // If the top bit of the output is demanded, demand everything from the
1104 // input. Otherwise, we demand all the input bits except NLZ top bits.
1105 uint64_t InDemandedBits = ~0ULL >> 64-BitWidth+NLZ;
1107 // Find information about known zero/one bits in the input.
1108 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1109 KnownZero2, KnownOne2, Depth+1))
1112 // If the RHS of the add has bits set that can't affect the input, reduce
1114 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1115 return UpdateValueUsesWith(I, I);
1117 // Avoid excess work.
1118 if (KnownZero2 == 0 && KnownOne2 == 0)
1121 // Turn it into OR if input bits are zero.
1122 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1124 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1126 InsertNewInstBefore(Or, *I);
1127 return UpdateValueUsesWith(I, Or);
1130 // We can say something about the output known-zero and known-one bits,
1131 // depending on potential carries from the input constant and the
1132 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1133 // bits set and the RHS constant is 0x01001, then we know we have a known
1134 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1136 // To compute this, we first compute the potential carry bits. These are
1137 // the bits which may be modified. I'm not aware of a better way to do
1139 uint64_t RHSVal = RHS->getZExtValue();
1141 bool CarryIn = false;
1142 uint64_t CarryBits = 0;
1143 uint64_t CurBit = 1;
1144 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1145 // Record the current carry in.
1146 if (CarryIn) CarryBits |= CurBit;
1150 // This bit has a carry out unless it is "zero + zero" or
1151 // "zero + anything" with no carry in.
1152 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1153 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1154 } else if (!CarryIn &&
1155 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1156 CarryOut = false; // 0 + anything has no carry out if no carry in.
1158 // Otherwise, we have to assume we have a carry out.
1162 // This stage's carry out becomes the next stage's carry-in.
1166 // Now that we know which bits have carries, compute the known-1/0 sets.
1168 // Bits are known one if they are known zero in one operand and one in the
1169 // other, and there is no input carry.
1170 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1172 // Bits are known zero if they are known zero in both operands and there
1173 // is no input carry.
1174 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1177 case Instruction::Shl:
1178 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1179 uint64_t ShiftAmt = SA->getZExtValue();
1180 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1181 KnownZero, KnownOne, Depth+1))
1183 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1184 KnownZero <<= ShiftAmt;
1185 KnownOne <<= ShiftAmt;
1186 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1189 case Instruction::LShr:
1190 // For a logical shift right
1191 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1192 unsigned ShiftAmt = SA->getZExtValue();
1194 // Compute the new bits that are at the top now.
1195 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1196 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1197 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1198 // Unsigned shift right.
1199 if (SimplifyDemandedBits(I->getOperand(0),
1200 (DemandedMask << ShiftAmt) & TypeMask,
1201 KnownZero, KnownOne, Depth+1))
1203 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1204 KnownZero &= TypeMask;
1205 KnownOne &= TypeMask;
1206 KnownZero >>= ShiftAmt;
1207 KnownOne >>= ShiftAmt;
1208 KnownZero |= HighBits; // high bits known zero.
1211 case Instruction::AShr:
1212 // If this is an arithmetic shift right and only the low-bit is set, we can
1213 // always convert this into a logical shr, even if the shift amount is
1214 // variable. The low bit of the shift cannot be an input sign bit unless
1215 // the shift amount is >= the size of the datatype, which is undefined.
1216 if (DemandedMask == 1) {
1217 // Perform the logical shift right.
1218 Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
1219 I->getOperand(1), I->getName());
1220 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1221 return UpdateValueUsesWith(I, NewVal);
1224 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1225 unsigned ShiftAmt = SA->getZExtValue();
1227 // Compute the new bits that are at the top now.
1228 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1229 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1230 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1231 // Signed shift right.
1232 if (SimplifyDemandedBits(I->getOperand(0),
1233 (DemandedMask << ShiftAmt) & TypeMask,
1234 KnownZero, KnownOne, Depth+1))
1236 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1237 KnownZero &= TypeMask;
1238 KnownOne &= TypeMask;
1239 KnownZero >>= ShiftAmt;
1240 KnownOne >>= ShiftAmt;
1242 // Handle the sign bits.
1243 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
1244 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1246 // If the input sign bit is known to be zero, or if none of the top bits
1247 // are demanded, turn this into an unsigned shift right.
1248 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1249 // Perform the logical shift right.
1250 Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
1252 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1253 return UpdateValueUsesWith(I, NewVal);
1254 } else if (KnownOne & SignBit) { // New bits are known one.
1255 KnownOne |= HighBits;
1261 // If the client is only demanding bits that we know, return the known
1263 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1264 return UpdateValueUsesWith(I, GetConstantInType(I->getType(), KnownOne));
1269 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1270 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1271 /// actually used by the caller. This method analyzes which elements of the
1272 /// operand are undef and returns that information in UndefElts.
1274 /// If the information about demanded elements can be used to simplify the
1275 /// operation, the operation is simplified, then the resultant value is
1276 /// returned. This returns null if no change was made.
1277 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1278 uint64_t &UndefElts,
1280 unsigned VWidth = cast<PackedType>(V->getType())->getNumElements();
1281 assert(VWidth <= 64 && "Vector too wide to analyze!");
1282 uint64_t EltMask = ~0ULL >> (64-VWidth);
1283 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1284 "Invalid DemandedElts!");
1286 if (isa<UndefValue>(V)) {
1287 // If the entire vector is undefined, just return this info.
1288 UndefElts = EltMask;
1290 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1291 UndefElts = EltMask;
1292 return UndefValue::get(V->getType());
1296 if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V)) {
1297 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1298 Constant *Undef = UndefValue::get(EltTy);
1300 std::vector<Constant*> Elts;
1301 for (unsigned i = 0; i != VWidth; ++i)
1302 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1303 Elts.push_back(Undef);
1304 UndefElts |= (1ULL << i);
1305 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1306 Elts.push_back(Undef);
1307 UndefElts |= (1ULL << i);
1308 } else { // Otherwise, defined.
1309 Elts.push_back(CP->getOperand(i));
1312 // If we changed the constant, return it.
1313 Constant *NewCP = ConstantPacked::get(Elts);
1314 return NewCP != CP ? NewCP : 0;
1315 } else if (isa<ConstantAggregateZero>(V)) {
1316 // Simplify the CAZ to a ConstantPacked where the non-demanded elements are
1318 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1319 Constant *Zero = Constant::getNullValue(EltTy);
1320 Constant *Undef = UndefValue::get(EltTy);
1321 std::vector<Constant*> Elts;
1322 for (unsigned i = 0; i != VWidth; ++i)
1323 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1324 UndefElts = DemandedElts ^ EltMask;
1325 return ConstantPacked::get(Elts);
1328 if (!V->hasOneUse()) { // Other users may use these bits.
1329 if (Depth != 0) { // Not at the root.
1330 // TODO: Just compute the UndefElts information recursively.
1334 } else if (Depth == 10) { // Limit search depth.
1338 Instruction *I = dyn_cast<Instruction>(V);
1339 if (!I) return false; // Only analyze instructions.
1341 bool MadeChange = false;
1342 uint64_t UndefElts2;
1344 switch (I->getOpcode()) {
1347 case Instruction::InsertElement: {
1348 // If this is a variable index, we don't know which element it overwrites.
1349 // demand exactly the same input as we produce.
1350 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1352 // Note that we can't propagate undef elt info, because we don't know
1353 // which elt is getting updated.
1354 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1355 UndefElts2, Depth+1);
1356 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1360 // If this is inserting an element that isn't demanded, remove this
1362 unsigned IdxNo = Idx->getZExtValue();
1363 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1364 return AddSoonDeadInstToWorklist(*I, 0);
1366 // Otherwise, the element inserted overwrites whatever was there, so the
1367 // input demanded set is simpler than the output set.
1368 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1369 DemandedElts & ~(1ULL << IdxNo),
1370 UndefElts, Depth+1);
1371 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1373 // The inserted element is defined.
1374 UndefElts |= 1ULL << IdxNo;
1378 case Instruction::And:
1379 case Instruction::Or:
1380 case Instruction::Xor:
1381 case Instruction::Add:
1382 case Instruction::Sub:
1383 case Instruction::Mul:
1384 // div/rem demand all inputs, because they don't want divide by zero.
1385 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1386 UndefElts, Depth+1);
1387 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1388 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1389 UndefElts2, Depth+1);
1390 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1392 // Output elements are undefined if both are undefined. Consider things
1393 // like undef&0. The result is known zero, not undef.
1394 UndefElts &= UndefElts2;
1397 case Instruction::Call: {
1398 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1400 switch (II->getIntrinsicID()) {
1403 // Binary vector operations that work column-wise. A dest element is a
1404 // function of the corresponding input elements from the two inputs.
1405 case Intrinsic::x86_sse_sub_ss:
1406 case Intrinsic::x86_sse_mul_ss:
1407 case Intrinsic::x86_sse_min_ss:
1408 case Intrinsic::x86_sse_max_ss:
1409 case Intrinsic::x86_sse2_sub_sd:
1410 case Intrinsic::x86_sse2_mul_sd:
1411 case Intrinsic::x86_sse2_min_sd:
1412 case Intrinsic::x86_sse2_max_sd:
1413 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1414 UndefElts, Depth+1);
1415 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1416 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1417 UndefElts2, Depth+1);
1418 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1420 // If only the low elt is demanded and this is a scalarizable intrinsic,
1421 // scalarize it now.
1422 if (DemandedElts == 1) {
1423 switch (II->getIntrinsicID()) {
1425 case Intrinsic::x86_sse_sub_ss:
1426 case Intrinsic::x86_sse_mul_ss:
1427 case Intrinsic::x86_sse2_sub_sd:
1428 case Intrinsic::x86_sse2_mul_sd:
1429 // TODO: Lower MIN/MAX/ABS/etc
1430 Value *LHS = II->getOperand(1);
1431 Value *RHS = II->getOperand(2);
1432 // Extract the element as scalars.
1433 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1434 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1436 switch (II->getIntrinsicID()) {
1437 default: assert(0 && "Case stmts out of sync!");
1438 case Intrinsic::x86_sse_sub_ss:
1439 case Intrinsic::x86_sse2_sub_sd:
1440 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1441 II->getName()), *II);
1443 case Intrinsic::x86_sse_mul_ss:
1444 case Intrinsic::x86_sse2_mul_sd:
1445 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1446 II->getName()), *II);
1451 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1453 InsertNewInstBefore(New, *II);
1454 AddSoonDeadInstToWorklist(*II, 0);
1459 // Output elements are undefined if both are undefined. Consider things
1460 // like undef&0. The result is known zero, not undef.
1461 UndefElts &= UndefElts2;
1467 return MadeChange ? I : 0;
1470 /// @returns true if the specified compare instruction is
1471 /// true when both operands are equal...
1472 /// @brief Determine if the ICmpInst returns true if both operands are equal
1473 static bool isTrueWhenEqual(ICmpInst &ICI) {
1474 ICmpInst::Predicate pred = ICI.getPredicate();
1475 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1476 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1477 pred == ICmpInst::ICMP_SLE;
1480 /// @returns true if the specified compare instruction is
1481 /// true when both operands are equal...
1482 /// @brief Determine if the FCmpInst returns true if both operands are equal
1483 static bool isTrueWhenEqual(FCmpInst &FCI) {
1484 FCmpInst::Predicate pred = FCI.getPredicate();
1485 return pred == FCmpInst::FCMP_OEQ || pred == FCmpInst::FCMP_UEQ ||
1486 pred == FCmpInst::FCMP_OGE || pred == FCmpInst::FCMP_UGE ||
1487 pred == FCmpInst::FCMP_OLE || pred == FCmpInst::FCMP_ULE;
1490 /// AssociativeOpt - Perform an optimization on an associative operator. This
1491 /// function is designed to check a chain of associative operators for a
1492 /// potential to apply a certain optimization. Since the optimization may be
1493 /// applicable if the expression was reassociated, this checks the chain, then
1494 /// reassociates the expression as necessary to expose the optimization
1495 /// opportunity. This makes use of a special Functor, which must define
1496 /// 'shouldApply' and 'apply' methods.
1498 template<typename Functor>
1499 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1500 unsigned Opcode = Root.getOpcode();
1501 Value *LHS = Root.getOperand(0);
1503 // Quick check, see if the immediate LHS matches...
1504 if (F.shouldApply(LHS))
1505 return F.apply(Root);
1507 // Otherwise, if the LHS is not of the same opcode as the root, return.
1508 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1509 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1510 // Should we apply this transform to the RHS?
1511 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1513 // If not to the RHS, check to see if we should apply to the LHS...
1514 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1515 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1519 // If the functor wants to apply the optimization to the RHS of LHSI,
1520 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1522 BasicBlock *BB = Root.getParent();
1524 // Now all of the instructions are in the current basic block, go ahead
1525 // and perform the reassociation.
1526 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1528 // First move the selected RHS to the LHS of the root...
1529 Root.setOperand(0, LHSI->getOperand(1));
1531 // Make what used to be the LHS of the root be the user of the root...
1532 Value *ExtraOperand = TmpLHSI->getOperand(1);
1533 if (&Root == TmpLHSI) {
1534 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1537 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1538 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1539 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1540 BasicBlock::iterator ARI = &Root; ++ARI;
1541 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1544 // Now propagate the ExtraOperand down the chain of instructions until we
1546 while (TmpLHSI != LHSI) {
1547 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1548 // Move the instruction to immediately before the chain we are
1549 // constructing to avoid breaking dominance properties.
1550 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1551 BB->getInstList().insert(ARI, NextLHSI);
1554 Value *NextOp = NextLHSI->getOperand(1);
1555 NextLHSI->setOperand(1, ExtraOperand);
1557 ExtraOperand = NextOp;
1560 // Now that the instructions are reassociated, have the functor perform
1561 // the transformation...
1562 return F.apply(Root);
1565 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1571 // AddRHS - Implements: X + X --> X << 1
1574 AddRHS(Value *rhs) : RHS(rhs) {}
1575 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1576 Instruction *apply(BinaryOperator &Add) const {
1577 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
1578 ConstantInt::get(Type::UByteTy, 1));
1582 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1584 struct AddMaskingAnd {
1586 AddMaskingAnd(Constant *c) : C2(c) {}
1587 bool shouldApply(Value *LHS) const {
1589 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1590 ConstantExpr::getAnd(C1, C2)->isNullValue();
1592 Instruction *apply(BinaryOperator &Add) const {
1593 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1597 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1599 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1600 if (Constant *SOC = dyn_cast<Constant>(SO))
1601 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1603 return IC->InsertNewInstBefore(CastInst::create(
1604 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1607 // Figure out if the constant is the left or the right argument.
1608 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1609 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1611 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1613 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1614 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1617 Value *Op0 = SO, *Op1 = ConstOperand;
1619 std::swap(Op0, Op1);
1621 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1622 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1623 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1624 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1625 SO->getName()+".cmp");
1626 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1627 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
1629 assert(0 && "Unknown binary instruction type!");
1632 return IC->InsertNewInstBefore(New, I);
1635 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1636 // constant as the other operand, try to fold the binary operator into the
1637 // select arguments. This also works for Cast instructions, which obviously do
1638 // not have a second operand.
1639 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1641 // Don't modify shared select instructions
1642 if (!SI->hasOneUse()) return 0;
1643 Value *TV = SI->getOperand(1);
1644 Value *FV = SI->getOperand(2);
1646 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1647 // Bool selects with constant operands can be folded to logical ops.
1648 if (SI->getType() == Type::BoolTy) return 0;
1650 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1651 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1653 return new SelectInst(SI->getCondition(), SelectTrueVal,
1660 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1661 /// node as operand #0, see if we can fold the instruction into the PHI (which
1662 /// is only possible if all operands to the PHI are constants).
1663 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1664 PHINode *PN = cast<PHINode>(I.getOperand(0));
1665 unsigned NumPHIValues = PN->getNumIncomingValues();
1666 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1668 // Check to see if all of the operands of the PHI are constants. If there is
1669 // one non-constant value, remember the BB it is. If there is more than one
1671 BasicBlock *NonConstBB = 0;
1672 for (unsigned i = 0; i != NumPHIValues; ++i)
1673 if (!isa<Constant>(PN->getIncomingValue(i))) {
1674 if (NonConstBB) return 0; // More than one non-const value.
1675 NonConstBB = PN->getIncomingBlock(i);
1677 // If the incoming non-constant value is in I's block, we have an infinite
1679 if (NonConstBB == I.getParent())
1683 // If there is exactly one non-constant value, we can insert a copy of the
1684 // operation in that block. However, if this is a critical edge, we would be
1685 // inserting the computation one some other paths (e.g. inside a loop). Only
1686 // do this if the pred block is unconditionally branching into the phi block.
1688 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1689 if (!BI || !BI->isUnconditional()) return 0;
1692 // Okay, we can do the transformation: create the new PHI node.
1693 PHINode *NewPN = new PHINode(I.getType(), I.getName());
1695 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1696 InsertNewInstBefore(NewPN, *PN);
1698 // Next, add all of the operands to the PHI.
1699 if (I.getNumOperands() == 2) {
1700 Constant *C = cast<Constant>(I.getOperand(1));
1701 for (unsigned i = 0; i != NumPHIValues; ++i) {
1703 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1704 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1705 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1707 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1709 assert(PN->getIncomingBlock(i) == NonConstBB);
1710 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1711 InV = BinaryOperator::create(BO->getOpcode(),
1712 PN->getIncomingValue(i), C, "phitmp",
1713 NonConstBB->getTerminator());
1714 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1715 InV = CmpInst::create(CI->getOpcode(),
1717 PN->getIncomingValue(i), C, "phitmp",
1718 NonConstBB->getTerminator());
1719 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1720 InV = new ShiftInst(SI->getOpcode(),
1721 PN->getIncomingValue(i), C, "phitmp",
1722 NonConstBB->getTerminator());
1724 assert(0 && "Unknown binop!");
1726 WorkList.push_back(cast<Instruction>(InV));
1728 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1731 CastInst *CI = cast<CastInst>(&I);
1732 const Type *RetTy = CI->getType();
1733 for (unsigned i = 0; i != NumPHIValues; ++i) {
1735 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1736 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1738 assert(PN->getIncomingBlock(i) == NonConstBB);
1739 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1740 I.getType(), "phitmp",
1741 NonConstBB->getTerminator());
1742 WorkList.push_back(cast<Instruction>(InV));
1744 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1747 return ReplaceInstUsesWith(I, NewPN);
1750 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1751 bool Changed = SimplifyCommutative(I);
1752 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1754 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1755 // X + undef -> undef
1756 if (isa<UndefValue>(RHS))
1757 return ReplaceInstUsesWith(I, RHS);
1760 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1761 if (RHSC->isNullValue())
1762 return ReplaceInstUsesWith(I, LHS);
1763 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1764 if (CFP->isExactlyValue(-0.0))
1765 return ReplaceInstUsesWith(I, LHS);
1768 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1769 // X + (signbit) --> X ^ signbit
1770 uint64_t Val = CI->getZExtValue();
1771 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1772 return BinaryOperator::createXor(LHS, RHS);
1774 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1775 // (X & 254)+1 -> (X&254)|1
1776 uint64_t KnownZero, KnownOne;
1777 if (!isa<PackedType>(I.getType()) &&
1778 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
1779 KnownZero, KnownOne))
1783 if (isa<PHINode>(LHS))
1784 if (Instruction *NV = FoldOpIntoPhi(I))
1787 ConstantInt *XorRHS = 0;
1789 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1790 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1791 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1792 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1794 uint64_t C0080Val = 1ULL << 31;
1795 int64_t CFF80Val = -C0080Val;
1798 if (TySizeBits > Size) {
1800 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1801 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1802 if (RHSSExt == CFF80Val) {
1803 if (XorRHS->getZExtValue() == C0080Val)
1805 } else if (RHSZExt == C0080Val) {
1806 if (XorRHS->getSExtValue() == CFF80Val)
1810 // This is a sign extend if the top bits are known zero.
1811 uint64_t Mask = ~0ULL;
1812 Mask <<= 64-(TySizeBits-Size);
1813 Mask &= XorLHS->getType()->getIntegralTypeMask();
1814 if (!MaskedValueIsZero(XorLHS, Mask))
1815 Size = 0; // Not a sign ext, but can't be any others either.
1822 } while (Size >= 8);
1825 const Type *MiddleType = 0;
1828 case 32: MiddleType = Type::IntTy; break;
1829 case 16: MiddleType = Type::ShortTy; break;
1830 case 8: MiddleType = Type::SByteTy; break;
1833 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1834 InsertNewInstBefore(NewTrunc, I);
1835 return new SExtInst(NewTrunc, I.getType());
1841 if (I.getType()->isInteger()) {
1842 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1844 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1845 if (RHSI->getOpcode() == Instruction::Sub)
1846 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1847 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1849 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1850 if (LHSI->getOpcode() == Instruction::Sub)
1851 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1852 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1857 if (Value *V = dyn_castNegVal(LHS))
1858 return BinaryOperator::createSub(RHS, V);
1861 if (!isa<Constant>(RHS))
1862 if (Value *V = dyn_castNegVal(RHS))
1863 return BinaryOperator::createSub(LHS, V);
1867 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1868 if (X == RHS) // X*C + X --> X * (C+1)
1869 return BinaryOperator::createMul(RHS, AddOne(C2));
1871 // X*C1 + X*C2 --> X * (C1+C2)
1873 if (X == dyn_castFoldableMul(RHS, C1))
1874 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1877 // X + X*C --> X * (C+1)
1878 if (dyn_castFoldableMul(RHS, C2) == LHS)
1879 return BinaryOperator::createMul(LHS, AddOne(C2));
1882 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1883 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1884 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2))) return R;
1886 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1888 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1889 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1890 return BinaryOperator::createSub(C, X);
1893 // (X & FF00) + xx00 -> (X+xx00) & FF00
1894 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1895 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1896 if (Anded == CRHS) {
1897 // See if all bits from the first bit set in the Add RHS up are included
1898 // in the mask. First, get the rightmost bit.
1899 uint64_t AddRHSV = CRHS->getZExtValue();
1901 // Form a mask of all bits from the lowest bit added through the top.
1902 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1903 AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
1905 // See if the and mask includes all of these bits.
1906 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1908 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1909 // Okay, the xform is safe. Insert the new add pronto.
1910 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1911 LHS->getName()), I);
1912 return BinaryOperator::createAnd(NewAdd, C2);
1917 // Try to fold constant add into select arguments.
1918 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1919 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1923 // add (cast *A to intptrtype) B ->
1924 // cast (GEP (cast *A to sbyte*) B) ->
1927 CastInst *CI = dyn_cast<CastInst>(LHS);
1930 CI = dyn_cast<CastInst>(RHS);
1933 if (CI && CI->getType()->isSized() &&
1934 (CI->getType()->getPrimitiveSize() ==
1935 TD->getIntPtrType()->getPrimitiveSize())
1936 && isa<PointerType>(CI->getOperand(0)->getType())) {
1937 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
1938 PointerType::get(Type::SByteTy), I);
1939 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1940 return new PtrToIntInst(I2, CI->getType());
1944 return Changed ? &I : 0;
1947 // isSignBit - Return true if the value represented by the constant only has the
1948 // highest order bit set.
1949 static bool isSignBit(ConstantInt *CI) {
1950 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1951 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1954 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1956 static Value *RemoveNoopCast(Value *V) {
1957 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1958 const Type *CTy = CI->getType();
1959 const Type *OpTy = CI->getOperand(0)->getType();
1960 if (CTy->isInteger() && OpTy->isInteger()) {
1961 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1962 return RemoveNoopCast(CI->getOperand(0));
1963 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1964 return RemoveNoopCast(CI->getOperand(0));
1969 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1970 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1972 if (Op0 == Op1) // sub X, X -> 0
1973 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1975 // If this is a 'B = x-(-A)', change to B = x+A...
1976 if (Value *V = dyn_castNegVal(Op1))
1977 return BinaryOperator::createAdd(Op0, V);
1979 if (isa<UndefValue>(Op0))
1980 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1981 if (isa<UndefValue>(Op1))
1982 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1984 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1985 // Replace (-1 - A) with (~A)...
1986 if (C->isAllOnesValue())
1987 return BinaryOperator::createNot(Op1);
1989 // C - ~X == X + (1+C)
1991 if (match(Op1, m_Not(m_Value(X))))
1992 return BinaryOperator::createAdd(X,
1993 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1994 // -((uint)X >> 31) -> ((int)X >> 31)
1995 // -((int)X >> 31) -> ((uint)X >> 31)
1996 if (C->isNullValue()) {
1997 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1998 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1999 if (SI->getOpcode() == Instruction::LShr) {
2000 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2001 // Check to see if we are shifting out everything but the sign bit.
2002 if (CU->getZExtValue() ==
2003 SI->getType()->getPrimitiveSizeInBits()-1) {
2004 // Ok, the transformation is safe. Insert AShr.
2005 return new ShiftInst(Instruction::AShr, SI->getOperand(0),
2010 else if (SI->getOpcode() == Instruction::AShr) {
2011 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2012 // Check to see if we are shifting out everything but the sign bit.
2013 if (CU->getZExtValue() ==
2014 SI->getType()->getPrimitiveSizeInBits()-1) {
2015 // Ok, the transformation is safe. Insert LShr.
2016 return new ShiftInst(Instruction::LShr, SI->getOperand(0),
2023 // Try to fold constant sub into select arguments.
2024 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2025 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2028 if (isa<PHINode>(Op0))
2029 if (Instruction *NV = FoldOpIntoPhi(I))
2033 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2034 if (Op1I->getOpcode() == Instruction::Add &&
2035 !Op0->getType()->isFPOrFPVector()) {
2036 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2037 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2038 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2039 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2040 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2041 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2042 // C1-(X+C2) --> (C1-C2)-X
2043 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2044 Op1I->getOperand(0));
2048 if (Op1I->hasOneUse()) {
2049 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2050 // is not used by anyone else...
2052 if (Op1I->getOpcode() == Instruction::Sub &&
2053 !Op1I->getType()->isFPOrFPVector()) {
2054 // Swap the two operands of the subexpr...
2055 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2056 Op1I->setOperand(0, IIOp1);
2057 Op1I->setOperand(1, IIOp0);
2059 // Create the new top level add instruction...
2060 return BinaryOperator::createAdd(Op0, Op1);
2063 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2065 if (Op1I->getOpcode() == Instruction::And &&
2066 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2067 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2070 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2071 return BinaryOperator::createAnd(Op0, NewNot);
2074 // 0 - (X sdiv C) -> (X sdiv -C)
2075 if (Op1I->getOpcode() == Instruction::SDiv)
2076 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2077 if (CSI->isNullValue())
2078 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2079 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2080 ConstantExpr::getNeg(DivRHS));
2082 // X - X*C --> X * (1-C)
2083 ConstantInt *C2 = 0;
2084 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2086 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2087 return BinaryOperator::createMul(Op0, CP1);
2092 if (!Op0->getType()->isFPOrFPVector())
2093 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2094 if (Op0I->getOpcode() == Instruction::Add) {
2095 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2096 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2097 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2098 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2099 } else if (Op0I->getOpcode() == Instruction::Sub) {
2100 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2101 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2105 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2106 if (X == Op1) { // X*C - X --> X * (C-1)
2107 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2108 return BinaryOperator::createMul(Op1, CP1);
2111 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2112 if (X == dyn_castFoldableMul(Op1, C2))
2113 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2118 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2119 /// really just returns true if the most significant (sign) bit is set.
2120 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2122 case ICmpInst::ICMP_SLT:
2123 // True if LHS s< RHS and RHS == 0
2124 return RHS->isNullValue();
2125 case ICmpInst::ICMP_SLE:
2126 // True if LHS s<= RHS and RHS == -1
2127 return RHS->isAllOnesValue();
2128 case ICmpInst::ICMP_UGE:
2129 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2130 return RHS->getZExtValue() == (1ULL <<
2131 (RHS->getType()->getPrimitiveSizeInBits()-1));
2132 case ICmpInst::ICMP_UGT:
2133 // True if LHS u> RHS and RHS == high-bit-mask - 1
2134 return RHS->getZExtValue() ==
2135 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2141 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2142 bool Changed = SimplifyCommutative(I);
2143 Value *Op0 = I.getOperand(0);
2145 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2146 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2148 // Simplify mul instructions with a constant RHS...
2149 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2150 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2152 // ((X << C1)*C2) == (X * (C2 << C1))
2153 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
2154 if (SI->getOpcode() == Instruction::Shl)
2155 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2156 return BinaryOperator::createMul(SI->getOperand(0),
2157 ConstantExpr::getShl(CI, ShOp));
2159 if (CI->isNullValue())
2160 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2161 if (CI->equalsInt(1)) // X * 1 == X
2162 return ReplaceInstUsesWith(I, Op0);
2163 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2164 return BinaryOperator::createNeg(Op0, I.getName());
2166 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2167 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2168 uint64_t C = Log2_64(Val);
2169 return new ShiftInst(Instruction::Shl, Op0,
2170 ConstantInt::get(Type::UByteTy, C));
2172 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2173 if (Op1F->isNullValue())
2174 return ReplaceInstUsesWith(I, Op1);
2176 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2177 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2178 if (Op1F->getValue() == 1.0)
2179 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2182 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2183 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2184 isa<ConstantInt>(Op0I->getOperand(1))) {
2185 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2186 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2188 InsertNewInstBefore(Add, I);
2189 Value *C1C2 = ConstantExpr::getMul(Op1,
2190 cast<Constant>(Op0I->getOperand(1)));
2191 return BinaryOperator::createAdd(Add, C1C2);
2195 // Try to fold constant mul into select arguments.
2196 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2197 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2200 if (isa<PHINode>(Op0))
2201 if (Instruction *NV = FoldOpIntoPhi(I))
2205 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2206 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2207 return BinaryOperator::createMul(Op0v, Op1v);
2209 // If one of the operands of the multiply is a cast from a boolean value, then
2210 // we know the bool is either zero or one, so this is a 'masking' multiply.
2211 // See if we can simplify things based on how the boolean was originally
2213 CastInst *BoolCast = 0;
2214 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2215 if (CI->getOperand(0)->getType() == Type::BoolTy)
2218 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2219 if (CI->getOperand(0)->getType() == Type::BoolTy)
2222 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2223 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2224 const Type *SCOpTy = SCIOp0->getType();
2226 // If the icmp is true iff the sign bit of X is set, then convert this
2227 // multiply into a shift/and combination.
2228 if (isa<ConstantInt>(SCIOp1) &&
2229 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2230 // Shift the X value right to turn it into "all signbits".
2231 Constant *Amt = ConstantInt::get(Type::UByteTy,
2232 SCOpTy->getPrimitiveSizeInBits()-1);
2234 InsertNewInstBefore(new ShiftInst(Instruction::AShr, SCIOp0, Amt,
2235 BoolCast->getOperand(0)->getName()+
2238 // If the multiply type is not the same as the source type, sign extend
2239 // or truncate to the multiply type.
2240 if (I.getType() != V->getType()) {
2241 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2242 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2243 Instruction::CastOps opcode =
2244 (SrcBits == DstBits ? Instruction::BitCast :
2245 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2246 V = InsertCastBefore(opcode, V, I.getType(), I);
2249 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2250 return BinaryOperator::createAnd(V, OtherOp);
2255 return Changed ? &I : 0;
2258 /// This function implements the transforms on div instructions that work
2259 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2260 /// used by the visitors to those instructions.
2261 /// @brief Transforms common to all three div instructions
2262 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2263 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2266 if (isa<UndefValue>(Op0))
2267 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2269 // X / undef -> undef
2270 if (isa<UndefValue>(Op1))
2271 return ReplaceInstUsesWith(I, Op1);
2273 // Handle cases involving: div X, (select Cond, Y, Z)
2274 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2275 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2276 // same basic block, then we replace the select with Y, and the condition
2277 // of the select with false (if the cond value is in the same BB). If the
2278 // select has uses other than the div, this allows them to be simplified
2279 // also. Note that div X, Y is just as good as div X, 0 (undef)
2280 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2281 if (ST->isNullValue()) {
2282 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2283 if (CondI && CondI->getParent() == I.getParent())
2284 UpdateValueUsesWith(CondI, ConstantBool::getFalse());
2285 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2286 I.setOperand(1, SI->getOperand(2));
2288 UpdateValueUsesWith(SI, SI->getOperand(2));
2292 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2293 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2294 if (ST->isNullValue()) {
2295 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2296 if (CondI && CondI->getParent() == I.getParent())
2297 UpdateValueUsesWith(CondI, ConstantBool::getTrue());
2298 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2299 I.setOperand(1, SI->getOperand(1));
2301 UpdateValueUsesWith(SI, SI->getOperand(1));
2309 /// This function implements the transforms common to both integer division
2310 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2311 /// division instructions.
2312 /// @brief Common integer divide transforms
2313 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2314 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2316 if (Instruction *Common = commonDivTransforms(I))
2319 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2321 if (RHS->equalsInt(1))
2322 return ReplaceInstUsesWith(I, Op0);
2324 // (X / C1) / C2 -> X / (C1*C2)
2325 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2326 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2327 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2328 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2329 ConstantExpr::getMul(RHS, LHSRHS));
2332 if (!RHS->isNullValue()) { // avoid X udiv 0
2333 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2334 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2336 if (isa<PHINode>(Op0))
2337 if (Instruction *NV = FoldOpIntoPhi(I))
2342 // 0 / X == 0, we don't need to preserve faults!
2343 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2344 if (LHS->equalsInt(0))
2345 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2350 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2351 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2353 // Handle the integer div common cases
2354 if (Instruction *Common = commonIDivTransforms(I))
2357 // X udiv C^2 -> X >> C
2358 // Check to see if this is an unsigned division with an exact power of 2,
2359 // if so, convert to a right shift.
2360 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2361 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2362 if (isPowerOf2_64(Val)) {
2363 uint64_t ShiftAmt = Log2_64(Val);
2364 return new ShiftInst(Instruction::LShr, Op0,
2365 ConstantInt::get(Type::UByteTy, ShiftAmt));
2369 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2370 if (ShiftInst *RHSI = dyn_cast<ShiftInst>(I.getOperand(1))) {
2371 if (RHSI->getOpcode() == Instruction::Shl &&
2372 isa<ConstantInt>(RHSI->getOperand(0))) {
2373 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2374 if (isPowerOf2_64(C1)) {
2375 Value *N = RHSI->getOperand(1);
2376 const Type *NTy = N->getType();
2377 if (uint64_t C2 = Log2_64(C1)) {
2378 Constant *C2V = ConstantInt::get(NTy, C2);
2379 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2381 return new ShiftInst(Instruction::LShr, Op0, N);
2386 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2387 // where C1&C2 are powers of two.
2388 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2389 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2390 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2391 if (!STO->isNullValue() && !STO->isNullValue()) {
2392 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2393 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2394 // Compute the shift amounts
2395 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2396 // Construct the "on true" case of the select
2397 Constant *TC = ConstantInt::get(Type::UByteTy, TSA);
2399 new ShiftInst(Instruction::LShr, Op0, TC, SI->getName()+".t");
2400 TSI = InsertNewInstBefore(TSI, I);
2402 // Construct the "on false" case of the select
2403 Constant *FC = ConstantInt::get(Type::UByteTy, FSA);
2405 new ShiftInst(Instruction::LShr, Op0, FC, SI->getName()+".f");
2406 FSI = InsertNewInstBefore(FSI, I);
2408 // construct the select instruction and return it.
2409 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2416 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2417 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2419 // Handle the integer div common cases
2420 if (Instruction *Common = commonIDivTransforms(I))
2423 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2425 if (RHS->isAllOnesValue())
2426 return BinaryOperator::createNeg(Op0);
2429 if (Value *LHSNeg = dyn_castNegVal(Op0))
2430 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2433 // If the sign bits of both operands are zero (i.e. we can prove they are
2434 // unsigned inputs), turn this into a udiv.
2435 if (I.getType()->isInteger()) {
2436 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2437 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2438 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2445 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2446 return commonDivTransforms(I);
2449 /// GetFactor - If we can prove that the specified value is at least a multiple
2450 /// of some factor, return that factor.
2451 static Constant *GetFactor(Value *V) {
2452 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2455 // Unless we can be tricky, we know this is a multiple of 1.
2456 Constant *Result = ConstantInt::get(V->getType(), 1);
2458 Instruction *I = dyn_cast<Instruction>(V);
2459 if (!I) return Result;
2461 if (I->getOpcode() == Instruction::Mul) {
2462 // Handle multiplies by a constant, etc.
2463 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2464 GetFactor(I->getOperand(1)));
2465 } else if (I->getOpcode() == Instruction::Shl) {
2466 // (X<<C) -> X * (1 << C)
2467 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2468 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2469 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2471 } else if (I->getOpcode() == Instruction::And) {
2472 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2473 // X & 0xFFF0 is known to be a multiple of 16.
2474 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2475 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2476 return ConstantExpr::getShl(Result,
2477 ConstantInt::get(Type::UByteTy, Zeros));
2479 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2480 // Only handle int->int casts.
2481 if (!CI->isIntegerCast())
2483 Value *Op = CI->getOperand(0);
2484 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2489 /// This function implements the transforms on rem instructions that work
2490 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2491 /// is used by the visitors to those instructions.
2492 /// @brief Transforms common to all three rem instructions
2493 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2494 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2496 // 0 % X == 0, we don't need to preserve faults!
2497 if (Constant *LHS = dyn_cast<Constant>(Op0))
2498 if (LHS->isNullValue())
2499 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2501 if (isa<UndefValue>(Op0)) // undef % X -> 0
2502 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2503 if (isa<UndefValue>(Op1))
2504 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2506 // Handle cases involving: rem X, (select Cond, Y, Z)
2507 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2508 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2509 // the same basic block, then we replace the select with Y, and the
2510 // condition of the select with false (if the cond value is in the same
2511 // BB). If the select has uses other than the div, this allows them to be
2513 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2514 if (ST->isNullValue()) {
2515 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2516 if (CondI && CondI->getParent() == I.getParent())
2517 UpdateValueUsesWith(CondI, ConstantBool::getFalse());
2518 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2519 I.setOperand(1, SI->getOperand(2));
2521 UpdateValueUsesWith(SI, SI->getOperand(2));
2524 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2525 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2526 if (ST->isNullValue()) {
2527 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2528 if (CondI && CondI->getParent() == I.getParent())
2529 UpdateValueUsesWith(CondI, ConstantBool::getTrue());
2530 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2531 I.setOperand(1, SI->getOperand(1));
2533 UpdateValueUsesWith(SI, SI->getOperand(1));
2541 /// This function implements the transforms common to both integer remainder
2542 /// instructions (urem and srem). It is called by the visitors to those integer
2543 /// remainder instructions.
2544 /// @brief Common integer remainder transforms
2545 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2546 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2548 if (Instruction *common = commonRemTransforms(I))
2551 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2552 // X % 0 == undef, we don't need to preserve faults!
2553 if (RHS->equalsInt(0))
2554 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2556 if (RHS->equalsInt(1)) // X % 1 == 0
2557 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2559 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2560 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2561 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2563 } else if (isa<PHINode>(Op0I)) {
2564 if (Instruction *NV = FoldOpIntoPhi(I))
2567 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2568 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2569 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2576 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2577 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2579 if (Instruction *common = commonIRemTransforms(I))
2582 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2583 // X urem C^2 -> X and C
2584 // Check to see if this is an unsigned remainder with an exact power of 2,
2585 // if so, convert to a bitwise and.
2586 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2587 if (isPowerOf2_64(C->getZExtValue()))
2588 return BinaryOperator::createAnd(Op0, SubOne(C));
2591 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2592 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2593 if (RHSI->getOpcode() == Instruction::Shl &&
2594 isa<ConstantInt>(RHSI->getOperand(0))) {
2595 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2596 if (isPowerOf2_64(C1)) {
2597 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2598 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2600 return BinaryOperator::createAnd(Op0, Add);
2605 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2606 // where C1&C2 are powers of two.
2607 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2608 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2609 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2610 // STO == 0 and SFO == 0 handled above.
2611 if (isPowerOf2_64(STO->getZExtValue()) &&
2612 isPowerOf2_64(SFO->getZExtValue())) {
2613 Value *TrueAnd = InsertNewInstBefore(
2614 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2615 Value *FalseAnd = InsertNewInstBefore(
2616 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2617 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2625 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2626 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2628 if (Instruction *common = commonIRemTransforms(I))
2631 if (Value *RHSNeg = dyn_castNegVal(Op1))
2632 if (!isa<ConstantInt>(RHSNeg) ||
2633 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2635 AddUsesToWorkList(I);
2636 I.setOperand(1, RHSNeg);
2640 // If the top bits of both operands are zero (i.e. we can prove they are
2641 // unsigned inputs), turn this into a urem.
2642 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2643 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2644 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2645 return BinaryOperator::createURem(Op0, Op1, I.getName());
2651 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2652 return commonRemTransforms(I);
2655 // isMaxValueMinusOne - return true if this is Max-1
2656 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2658 // Calculate 0111111111..11111
2659 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2660 int64_t Val = INT64_MAX; // All ones
2661 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2662 return C->getSExtValue() == Val-1;
2664 return C->getZExtValue() == C->getType()->getIntegralTypeMask()-1;
2667 // isMinValuePlusOne - return true if this is Min+1
2668 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2670 // Calculate 1111111111000000000000
2671 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2672 int64_t Val = -1; // All ones
2673 Val <<= TypeBits-1; // Shift over to the right spot
2674 return C->getSExtValue() == Val+1;
2676 return C->getZExtValue() == 1; // unsigned
2679 // isOneBitSet - Return true if there is exactly one bit set in the specified
2681 static bool isOneBitSet(const ConstantInt *CI) {
2682 uint64_t V = CI->getZExtValue();
2683 return V && (V & (V-1)) == 0;
2686 #if 0 // Currently unused
2687 // isLowOnes - Return true if the constant is of the form 0+1+.
2688 static bool isLowOnes(const ConstantInt *CI) {
2689 uint64_t V = CI->getZExtValue();
2691 // There won't be bits set in parts that the type doesn't contain.
2692 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2694 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2695 return U && V && (U & V) == 0;
2699 // isHighOnes - Return true if the constant is of the form 1+0+.
2700 // This is the same as lowones(~X).
2701 static bool isHighOnes(const ConstantInt *CI) {
2702 uint64_t V = ~CI->getZExtValue();
2703 if (~V == 0) return false; // 0's does not match "1+"
2705 // There won't be bits set in parts that the type doesn't contain.
2706 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2708 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2709 return U && V && (U & V) == 0;
2712 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2713 /// are carefully arranged to allow folding of expressions such as:
2715 /// (A < B) | (A > B) --> (A != B)
2717 /// Note that this is only valid if the first and second predicates have the
2718 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2720 /// Three bits are used to represent the condition, as follows:
2725 /// <=> Value Definition
2726 /// 000 0 Always false
2733 /// 111 7 Always true
2735 static unsigned getICmpCode(const ICmpInst *ICI) {
2736 switch (ICI->getPredicate()) {
2738 case ICmpInst::ICMP_UGT: return 1; // 001
2739 case ICmpInst::ICMP_SGT: return 1; // 001
2740 case ICmpInst::ICMP_EQ: return 2; // 010
2741 case ICmpInst::ICMP_UGE: return 3; // 011
2742 case ICmpInst::ICMP_SGE: return 3; // 011
2743 case ICmpInst::ICMP_ULT: return 4; // 100
2744 case ICmpInst::ICMP_SLT: return 4; // 100
2745 case ICmpInst::ICMP_NE: return 5; // 101
2746 case ICmpInst::ICMP_ULE: return 6; // 110
2747 case ICmpInst::ICMP_SLE: return 6; // 110
2750 assert(0 && "Invalid ICmp predicate!");
2755 /// getICmpValue - This is the complement of getICmpCode, which turns an
2756 /// opcode and two operands into either a constant true or false, or a brand
2757 /// new /// ICmp instruction. The sign is passed in to determine which kind
2758 /// of predicate to use in new icmp instructions.
2759 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2761 default: assert(0 && "Illegal ICmp code!");
2762 case 0: return ConstantBool::getFalse();
2765 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2767 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2768 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2771 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2773 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2776 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2778 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2779 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2782 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2784 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2785 case 7: return ConstantBool::getTrue();
2789 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2790 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2791 (ICmpInst::isSignedPredicate(p1) &&
2792 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2793 (ICmpInst::isSignedPredicate(p2) &&
2794 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2798 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2799 struct FoldICmpLogical {
2802 ICmpInst::Predicate pred;
2803 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2804 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2805 pred(ICI->getPredicate()) {}
2806 bool shouldApply(Value *V) const {
2807 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2808 if (PredicatesFoldable(pred, ICI->getPredicate()))
2809 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2810 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2813 Instruction *apply(Instruction &Log) const {
2814 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2815 if (ICI->getOperand(0) != LHS) {
2816 assert(ICI->getOperand(1) == LHS);
2817 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2820 unsigned LHSCode = getICmpCode(ICI);
2821 unsigned RHSCode = getICmpCode(cast<ICmpInst>(Log.getOperand(1)));
2823 switch (Log.getOpcode()) {
2824 case Instruction::And: Code = LHSCode & RHSCode; break;
2825 case Instruction::Or: Code = LHSCode | RHSCode; break;
2826 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2827 default: assert(0 && "Illegal logical opcode!"); return 0;
2830 Value *RV = getICmpValue(ICmpInst::isSignedPredicate(pred), Code, LHS, RHS);
2831 if (Instruction *I = dyn_cast<Instruction>(RV))
2833 // Otherwise, it's a constant boolean value...
2834 return IC.ReplaceInstUsesWith(Log, RV);
2837 } // end anonymous namespace
2839 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2840 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2841 // guaranteed to be either a shift instruction or a binary operator.
2842 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2843 ConstantIntegral *OpRHS,
2844 ConstantIntegral *AndRHS,
2845 BinaryOperator &TheAnd) {
2846 Value *X = Op->getOperand(0);
2847 Constant *Together = 0;
2848 if (!isa<ShiftInst>(Op))
2849 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2851 switch (Op->getOpcode()) {
2852 case Instruction::Xor:
2853 if (Op->hasOneUse()) {
2854 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2855 std::string OpName = Op->getName(); Op->setName("");
2856 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
2857 InsertNewInstBefore(And, TheAnd);
2858 return BinaryOperator::createXor(And, Together);
2861 case Instruction::Or:
2862 if (Together == AndRHS) // (X | C) & C --> C
2863 return ReplaceInstUsesWith(TheAnd, AndRHS);
2865 if (Op->hasOneUse() && Together != OpRHS) {
2866 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2867 std::string Op0Name = Op->getName(); Op->setName("");
2868 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
2869 InsertNewInstBefore(Or, TheAnd);
2870 return BinaryOperator::createAnd(Or, AndRHS);
2873 case Instruction::Add:
2874 if (Op->hasOneUse()) {
2875 // Adding a one to a single bit bit-field should be turned into an XOR
2876 // of the bit. First thing to check is to see if this AND is with a
2877 // single bit constant.
2878 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2880 // Clear bits that are not part of the constant.
2881 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
2883 // If there is only one bit set...
2884 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2885 // Ok, at this point, we know that we are masking the result of the
2886 // ADD down to exactly one bit. If the constant we are adding has
2887 // no bits set below this bit, then we can eliminate the ADD.
2888 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2890 // Check to see if any bits below the one bit set in AndRHSV are set.
2891 if ((AddRHS & (AndRHSV-1)) == 0) {
2892 // If not, the only thing that can effect the output of the AND is
2893 // the bit specified by AndRHSV. If that bit is set, the effect of
2894 // the XOR is to toggle the bit. If it is clear, then the ADD has
2896 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2897 TheAnd.setOperand(0, X);
2900 std::string Name = Op->getName(); Op->setName("");
2901 // Pull the XOR out of the AND.
2902 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
2903 InsertNewInstBefore(NewAnd, TheAnd);
2904 return BinaryOperator::createXor(NewAnd, AndRHS);
2911 case Instruction::Shl: {
2912 // We know that the AND will not produce any of the bits shifted in, so if
2913 // the anded constant includes them, clear them now!
2915 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2916 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2917 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2919 if (CI == ShlMask) { // Masking out bits that the shift already masks
2920 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2921 } else if (CI != AndRHS) { // Reducing bits set in and.
2922 TheAnd.setOperand(1, CI);
2927 case Instruction::LShr:
2929 // We know that the AND will not produce any of the bits shifted in, so if
2930 // the anded constant includes them, clear them now! This only applies to
2931 // unsigned shifts, because a signed shr may bring in set bits!
2933 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2934 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2935 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2937 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2938 return ReplaceInstUsesWith(TheAnd, Op);
2939 } else if (CI != AndRHS) {
2940 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2945 case Instruction::AShr:
2947 // See if this is shifting in some sign extension, then masking it out
2949 if (Op->hasOneUse()) {
2950 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2951 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2952 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
2953 if (C == AndRHS) { // Masking out bits shifted in.
2954 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
2955 // Make the argument unsigned.
2956 Value *ShVal = Op->getOperand(0);
2957 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::LShr, ShVal,
2958 OpRHS, Op->getName()), TheAnd);
2959 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
2968 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2969 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2970 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
2971 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
2972 /// insert new instructions.
2973 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2974 bool isSigned, bool Inside,
2976 assert(cast<ConstantBool>(ConstantExpr::getICmp((isSigned ?
2977 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getValue() &&
2978 "Lo is not <= Hi in range emission code!");
2981 if (Lo == Hi) // Trivially false.
2982 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
2984 // V >= Min && V < Hi --> V < Hi
2985 if (cast<ConstantIntegral>(Lo)->isMinValue(isSigned)) {
2986 ICmpInst::Predicate pred = (isSigned ?
2987 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
2988 return new ICmpInst(pred, V, Hi);
2991 // Emit V-Lo <u Hi-Lo
2992 Constant *NegLo = ConstantExpr::getNeg(Lo);
2993 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
2994 InsertNewInstBefore(Add, IB);
2995 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
2996 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
2999 if (Lo == Hi) // Trivially true.
3000 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3002 // V < Min || V >= Hi ->'V > Hi-1'
3003 Hi = SubOne(cast<ConstantInt>(Hi));
3004 if (cast<ConstantIntegral>(Lo)->isMinValue(isSigned)) {
3005 ICmpInst::Predicate pred = (isSigned ?
3006 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3007 return new ICmpInst(pred, V, Hi);
3010 // Emit V-Lo > Hi-1-Lo
3011 Constant *NegLo = ConstantExpr::getNeg(Lo);
3012 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3013 InsertNewInstBefore(Add, IB);
3014 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3015 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3018 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3019 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3020 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3021 // not, since all 1s are not contiguous.
3022 static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
3023 uint64_t V = Val->getZExtValue();
3024 if (!isShiftedMask_64(V)) return false;
3026 // look for the first zero bit after the run of ones
3027 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
3028 // look for the first non-zero bit
3029 ME = 64-CountLeadingZeros_64(V);
3035 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3036 /// where isSub determines whether the operator is a sub. If we can fold one of
3037 /// the following xforms:
3039 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3040 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3041 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3043 /// return (A +/- B).
3045 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3046 ConstantIntegral *Mask, bool isSub,
3048 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3049 if (!LHSI || LHSI->getNumOperands() != 2 ||
3050 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3052 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3054 switch (LHSI->getOpcode()) {
3056 case Instruction::And:
3057 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3058 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3059 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
3062 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3063 // part, we don't need any explicit masks to take them out of A. If that
3064 // is all N is, ignore it.
3066 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3067 uint64_t Mask = RHS->getType()->getIntegralTypeMask();
3069 if (MaskedValueIsZero(RHS, Mask))
3074 case Instruction::Or:
3075 case Instruction::Xor:
3076 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3077 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
3078 ConstantExpr::getAnd(N, Mask)->isNullValue())
3085 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3087 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3088 return InsertNewInstBefore(New, I);
3091 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3092 bool Changed = SimplifyCommutative(I);
3093 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3095 if (isa<UndefValue>(Op1)) // X & undef -> 0
3096 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3100 return ReplaceInstUsesWith(I, Op1);
3102 // See if we can simplify any instructions used by the instruction whose sole
3103 // purpose is to compute bits we don't care about.
3104 uint64_t KnownZero, KnownOne;
3105 if (!isa<PackedType>(I.getType()) &&
3106 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3107 KnownZero, KnownOne))
3110 if (ConstantIntegral *AndRHS = dyn_cast<ConstantIntegral>(Op1)) {
3111 uint64_t AndRHSMask = AndRHS->getZExtValue();
3112 uint64_t TypeMask = Op0->getType()->getIntegralTypeMask();
3113 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3115 // Optimize a variety of ((val OP C1) & C2) combinations...
3116 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
3117 Instruction *Op0I = cast<Instruction>(Op0);
3118 Value *Op0LHS = Op0I->getOperand(0);
3119 Value *Op0RHS = Op0I->getOperand(1);
3120 switch (Op0I->getOpcode()) {
3121 case Instruction::Xor:
3122 case Instruction::Or:
3123 // If the mask is only needed on one incoming arm, push it up.
3124 if (Op0I->hasOneUse()) {
3125 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3126 // Not masking anything out for the LHS, move to RHS.
3127 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3128 Op0RHS->getName()+".masked");
3129 InsertNewInstBefore(NewRHS, I);
3130 return BinaryOperator::create(
3131 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3133 if (!isa<Constant>(Op0RHS) &&
3134 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3135 // Not masking anything out for the RHS, move to LHS.
3136 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3137 Op0LHS->getName()+".masked");
3138 InsertNewInstBefore(NewLHS, I);
3139 return BinaryOperator::create(
3140 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3145 case Instruction::Add:
3146 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3147 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3148 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3149 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3150 return BinaryOperator::createAnd(V, AndRHS);
3151 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3152 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3155 case Instruction::Sub:
3156 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3157 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3158 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3159 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3160 return BinaryOperator::createAnd(V, AndRHS);
3164 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3165 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3167 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3168 // If this is an integer truncation or change from signed-to-unsigned, and
3169 // if the source is an and/or with immediate, transform it. This
3170 // frequently occurs for bitfield accesses.
3171 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3172 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3173 CastOp->getNumOperands() == 2)
3174 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3175 if (CastOp->getOpcode() == Instruction::And) {
3176 // Change: and (cast (and X, C1) to T), C2
3177 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3178 // This will fold the two constants together, which may allow
3179 // other simplifications.
3180 Instruction *NewCast = CastInst::createTruncOrBitCast(
3181 CastOp->getOperand(0), I.getType(),
3182 CastOp->getName()+".shrunk");
3183 NewCast = InsertNewInstBefore(NewCast, I);
3184 // trunc_or_bitcast(C1)&C2
3185 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3186 C3 = ConstantExpr::getAnd(C3, AndRHS);
3187 return BinaryOperator::createAnd(NewCast, C3);
3188 } else if (CastOp->getOpcode() == Instruction::Or) {
3189 // Change: and (cast (or X, C1) to T), C2
3190 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3191 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3192 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3193 return ReplaceInstUsesWith(I, AndRHS);
3198 // Try to fold constant and into select arguments.
3199 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3200 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3202 if (isa<PHINode>(Op0))
3203 if (Instruction *NV = FoldOpIntoPhi(I))
3207 Value *Op0NotVal = dyn_castNotVal(Op0);
3208 Value *Op1NotVal = dyn_castNotVal(Op1);
3210 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3211 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3213 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3214 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3215 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3216 I.getName()+".demorgan");
3217 InsertNewInstBefore(Or, I);
3218 return BinaryOperator::createNot(Or);
3222 Value *A = 0, *B = 0;
3223 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3224 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3225 return ReplaceInstUsesWith(I, Op1);
3226 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3227 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3228 return ReplaceInstUsesWith(I, Op0);
3230 if (Op0->hasOneUse() &&
3231 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3232 if (A == Op1) { // (A^B)&A -> A&(A^B)
3233 I.swapOperands(); // Simplify below
3234 std::swap(Op0, Op1);
3235 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3236 cast<BinaryOperator>(Op0)->swapOperands();
3237 I.swapOperands(); // Simplify below
3238 std::swap(Op0, Op1);
3241 if (Op1->hasOneUse() &&
3242 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3243 if (B == Op0) { // B&(A^B) -> B&(B^A)
3244 cast<BinaryOperator>(Op1)->swapOperands();
3247 if (A == Op0) { // A&(A^B) -> A & ~B
3248 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3249 InsertNewInstBefore(NotB, I);
3250 return BinaryOperator::createAnd(A, NotB);
3255 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3256 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3257 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3260 Value *LHSVal, *RHSVal;
3261 ConstantInt *LHSCst, *RHSCst;
3262 ICmpInst::Predicate LHSCC, RHSCC;
3263 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3264 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3265 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3266 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3267 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3268 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3269 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3270 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3271 // Ensure that the larger constant is on the RHS.
3272 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3273 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3274 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3275 ICmpInst *LHS = cast<ICmpInst>(Op0);
3276 if (cast<ConstantBool>(Cmp)->getValue()) {
3277 std::swap(LHS, RHS);
3278 std::swap(LHSCst, RHSCst);
3279 std::swap(LHSCC, RHSCC);
3282 // At this point, we know we have have two icmp instructions
3283 // comparing a value against two constants and and'ing the result
3284 // together. Because of the above check, we know that we only have
3285 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3286 // (from the FoldICmpLogical check above), that the two constants
3287 // are not equal and that the larger constant is on the RHS
3288 assert(LHSCst != RHSCst && "Compares not folded above?");
3291 default: assert(0 && "Unknown integer condition code!");
3292 case ICmpInst::ICMP_EQ:
3294 default: assert(0 && "Unknown integer condition code!");
3295 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3296 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3297 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3298 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
3299 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3300 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3301 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3302 return ReplaceInstUsesWith(I, LHS);
3304 case ICmpInst::ICMP_NE:
3306 default: assert(0 && "Unknown integer condition code!");
3307 case ICmpInst::ICMP_ULT:
3308 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3309 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3310 break; // (X != 13 & X u< 15) -> no change
3311 case ICmpInst::ICMP_SLT:
3312 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3313 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3314 break; // (X != 13 & X s< 15) -> no change
3315 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3316 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3317 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3318 return ReplaceInstUsesWith(I, RHS);
3319 case ICmpInst::ICMP_NE:
3320 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3321 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3322 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3323 LHSVal->getName()+".off");
3324 InsertNewInstBefore(Add, I);
3325 const Type *UnsType = Add->getType()->getUnsignedVersion();
3326 Value *OffsetVal = InsertCastBefore(Instruction::BitCast, Add,
3328 AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
3329 AddCST = ConstantExpr::getBitCast(AddCST, UnsType);
3330 return new ICmpInst(ICmpInst::ICMP_UGT, OffsetVal, AddCST);
3332 break; // (X != 13 & X != 15) -> no change
3335 case ICmpInst::ICMP_ULT:
3337 default: assert(0 && "Unknown integer condition code!");
3338 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3339 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3340 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
3341 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3343 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3344 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3345 return ReplaceInstUsesWith(I, LHS);
3346 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3350 case ICmpInst::ICMP_SLT:
3352 default: assert(0 && "Unknown integer condition code!");
3353 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3354 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3355 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
3356 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3358 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3359 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3360 return ReplaceInstUsesWith(I, LHS);
3361 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3365 case ICmpInst::ICMP_UGT:
3367 default: assert(0 && "Unknown integer condition code!");
3368 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3369 return ReplaceInstUsesWith(I, LHS);
3370 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3371 return ReplaceInstUsesWith(I, RHS);
3372 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3374 case ICmpInst::ICMP_NE:
3375 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3376 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3377 break; // (X u> 13 & X != 15) -> no change
3378 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3379 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3381 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3385 case ICmpInst::ICMP_SGT:
3387 default: assert(0 && "Unknown integer condition code!");
3388 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3389 return ReplaceInstUsesWith(I, LHS);
3390 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3391 return ReplaceInstUsesWith(I, RHS);
3392 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3394 case ICmpInst::ICMP_NE:
3395 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3396 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3397 break; // (X s> 13 & X != 15) -> no change
3398 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3399 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3401 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3409 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3410 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3411 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3412 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3413 const Type *SrcTy = Op0C->getOperand(0)->getType();
3414 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3415 // Only do this if the casts both really cause code to be generated.
3416 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3418 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3420 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3421 Op1C->getOperand(0),
3423 InsertNewInstBefore(NewOp, I);
3424 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3428 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3429 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3430 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3431 if (SI0->getOpcode() == SI1->getOpcode() &&
3432 SI0->getOperand(1) == SI1->getOperand(1) &&
3433 (SI0->hasOneUse() || SI1->hasOneUse())) {
3434 Instruction *NewOp =
3435 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3437 SI0->getName()), I);
3438 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3442 return Changed ? &I : 0;
3445 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3446 /// in the result. If it does, and if the specified byte hasn't been filled in
3447 /// yet, fill it in and return false.
3448 static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
3449 Instruction *I = dyn_cast<Instruction>(V);
3450 if (I == 0) return true;
3452 // If this is an or instruction, it is an inner node of the bswap.
3453 if (I->getOpcode() == Instruction::Or)
3454 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3455 CollectBSwapParts(I->getOperand(1), ByteValues);
3457 // If this is a shift by a constant int, and it is "24", then its operand
3458 // defines a byte. We only handle unsigned types here.
3459 if (isa<ShiftInst>(I) && isa<ConstantInt>(I->getOperand(1))) {
3460 // Not shifting the entire input by N-1 bytes?
3461 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3462 8*(ByteValues.size()-1))
3466 if (I->getOpcode() == Instruction::Shl) {
3467 // X << 24 defines the top byte with the lowest of the input bytes.
3468 DestNo = ByteValues.size()-1;
3470 // X >>u 24 defines the low byte with the highest of the input bytes.
3474 // If the destination byte value is already defined, the values are or'd
3475 // together, which isn't a bswap (unless it's an or of the same bits).
3476 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3478 ByteValues[DestNo] = I->getOperand(0);
3482 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3484 Value *Shift = 0, *ShiftLHS = 0;
3485 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3486 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3487 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3489 Instruction *SI = cast<Instruction>(Shift);
3491 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3492 if (ShiftAmt->getZExtValue() & 7 ||
3493 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3496 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3498 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3499 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3501 // Unknown mask for bswap.
3502 if (DestByte == ByteValues.size()) return true;
3504 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3506 if (SI->getOpcode() == Instruction::Shl)
3507 SrcByte = DestByte - ShiftBytes;
3509 SrcByte = DestByte + ShiftBytes;
3511 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3512 if (SrcByte != ByteValues.size()-DestByte-1)
3515 // If the destination byte value is already defined, the values are or'd
3516 // together, which isn't a bswap (unless it's an or of the same bits).
3517 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3519 ByteValues[DestByte] = SI->getOperand(0);
3523 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3524 /// If so, insert the new bswap intrinsic and return it.
3525 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3526 // We can only handle bswap of unsigned integers, and cannot bswap one byte.
3527 if (!I.getType()->isUnsigned() || I.getType() == Type::UByteTy)
3530 /// ByteValues - For each byte of the result, we keep track of which value
3531 /// defines each byte.
3532 std::vector<Value*> ByteValues;
3533 ByteValues.resize(I.getType()->getPrimitiveSize());
3535 // Try to find all the pieces corresponding to the bswap.
3536 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3537 CollectBSwapParts(I.getOperand(1), ByteValues))
3540 // Check to see if all of the bytes come from the same value.
3541 Value *V = ByteValues[0];
3542 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3544 // Check to make sure that all of the bytes come from the same value.
3545 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3546 if (ByteValues[i] != V)
3549 // If they do then *success* we can turn this into a bswap. Figure out what
3550 // bswap to make it into.
3551 Module *M = I.getParent()->getParent()->getParent();
3552 const char *FnName = 0;
3553 if (I.getType() == Type::UShortTy)
3554 FnName = "llvm.bswap.i16";
3555 else if (I.getType() == Type::UIntTy)
3556 FnName = "llvm.bswap.i32";
3557 else if (I.getType() == Type::ULongTy)
3558 FnName = "llvm.bswap.i64";
3560 assert(0 && "Unknown integer type!");
3561 Function *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3563 return new CallInst(F, V);
3567 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3568 bool Changed = SimplifyCommutative(I);
3569 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3571 if (isa<UndefValue>(Op1))
3572 return ReplaceInstUsesWith(I, // X | undef -> -1
3573 ConstantIntegral::getAllOnesValue(I.getType()));
3577 return ReplaceInstUsesWith(I, Op0);
3579 // See if we can simplify any instructions used by the instruction whose sole
3580 // purpose is to compute bits we don't care about.
3581 uint64_t KnownZero, KnownOne;
3582 if (!isa<PackedType>(I.getType()) &&
3583 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3584 KnownZero, KnownOne))
3588 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3589 ConstantInt *C1 = 0; Value *X = 0;
3590 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3591 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3592 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
3594 InsertNewInstBefore(Or, I);
3595 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3598 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3599 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3600 std::string Op0Name = Op0->getName(); Op0->setName("");
3601 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
3602 InsertNewInstBefore(Or, I);
3603 return BinaryOperator::createXor(Or,
3604 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3607 // Try to fold constant and into select arguments.
3608 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3609 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3611 if (isa<PHINode>(Op0))
3612 if (Instruction *NV = FoldOpIntoPhi(I))
3616 Value *A = 0, *B = 0;
3617 ConstantInt *C1 = 0, *C2 = 0;
3619 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3620 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3621 return ReplaceInstUsesWith(I, Op1);
3622 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3623 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3624 return ReplaceInstUsesWith(I, Op0);
3626 // (A | B) | C and A | (B | C) -> bswap if possible.
3627 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3628 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3629 match(Op1, m_Or(m_Value(), m_Value())) ||
3630 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3631 match(Op1, m_Shift(m_Value(), m_Value())))) {
3632 if (Instruction *BSwap = MatchBSwap(I))
3636 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3637 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3638 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3639 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
3641 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3644 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3645 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3646 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3647 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
3649 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3652 // (A & C1)|(B & C2)
3653 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3654 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3656 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3657 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3660 // If we have: ((V + N) & C1) | (V & C2)
3661 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3662 // replace with V+N.
3663 if (C1 == ConstantExpr::getNot(C2)) {
3664 Value *V1 = 0, *V2 = 0;
3665 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3666 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3667 // Add commutes, try both ways.
3668 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3669 return ReplaceInstUsesWith(I, A);
3670 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3671 return ReplaceInstUsesWith(I, A);
3673 // Or commutes, try both ways.
3674 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3675 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3676 // Add commutes, try both ways.
3677 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3678 return ReplaceInstUsesWith(I, B);
3679 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3680 return ReplaceInstUsesWith(I, B);
3685 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3686 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3687 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3688 if (SI0->getOpcode() == SI1->getOpcode() &&
3689 SI0->getOperand(1) == SI1->getOperand(1) &&
3690 (SI0->hasOneUse() || SI1->hasOneUse())) {
3691 Instruction *NewOp =
3692 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3694 SI0->getName()), I);
3695 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3699 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3700 if (A == Op1) // ~A | A == -1
3701 return ReplaceInstUsesWith(I,
3702 ConstantIntegral::getAllOnesValue(I.getType()));
3706 // Note, A is still live here!
3707 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3709 return ReplaceInstUsesWith(I,
3710 ConstantIntegral::getAllOnesValue(I.getType()));
3712 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3713 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3714 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3715 I.getName()+".demorgan"), I);
3716 return BinaryOperator::createNot(And);
3720 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3721 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3722 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3725 Value *LHSVal, *RHSVal;
3726 ConstantInt *LHSCst, *RHSCst;
3727 ICmpInst::Predicate LHSCC, RHSCC;
3728 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3729 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3730 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3731 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3732 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3733 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3734 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3735 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3736 // Ensure that the larger constant is on the RHS.
3737 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3738 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3739 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3740 ICmpInst *LHS = cast<ICmpInst>(Op0);
3741 if (cast<ConstantBool>(Cmp)->getValue()) {
3742 std::swap(LHS, RHS);
3743 std::swap(LHSCst, RHSCst);
3744 std::swap(LHSCC, RHSCC);
3747 // At this point, we know we have have two icmp instructions
3748 // comparing a value against two constants and or'ing the result
3749 // together. Because of the above check, we know that we only have
3750 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3751 // FoldICmpLogical check above), that the two constants are not
3753 assert(LHSCst != RHSCst && "Compares not folded above?");
3756 default: assert(0 && "Unknown integer condition code!");
3757 case ICmpInst::ICMP_EQ:
3759 default: assert(0 && "Unknown integer condition code!");
3760 case ICmpInst::ICMP_EQ:
3761 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3762 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3763 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3764 LHSVal->getName()+".off");
3765 InsertNewInstBefore(Add, I);
3766 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3767 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3769 break; // (X == 13 | X == 15) -> no change
3770 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3771 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3773 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3774 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3775 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3776 return ReplaceInstUsesWith(I, RHS);
3779 case ICmpInst::ICMP_NE:
3781 default: assert(0 && "Unknown integer condition code!");
3782 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3783 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3784 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3785 return ReplaceInstUsesWith(I, LHS);
3786 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3787 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3788 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3789 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
3792 case ICmpInst::ICMP_ULT:
3794 default: assert(0 && "Unknown integer condition code!");
3795 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3797 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3798 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3800 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3802 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3803 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3804 return ReplaceInstUsesWith(I, RHS);
3805 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
3809 case ICmpInst::ICMP_SLT:
3811 default: assert(0 && "Unknown integer condition code!");
3812 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
3814 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
3815 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
3817 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
3819 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
3820 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
3821 return ReplaceInstUsesWith(I, RHS);
3822 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
3826 case ICmpInst::ICMP_UGT:
3828 default: assert(0 && "Unknown integer condition code!");
3829 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
3830 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
3831 return ReplaceInstUsesWith(I, LHS);
3832 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
3834 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
3835 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
3836 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
3837 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
3841 case ICmpInst::ICMP_SGT:
3843 default: assert(0 && "Unknown integer condition code!");
3844 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
3845 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
3846 return ReplaceInstUsesWith(I, LHS);
3847 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
3849 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
3850 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
3851 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
3852 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
3860 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3861 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3862 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3863 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3864 const Type *SrcTy = Op0C->getOperand(0)->getType();
3865 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3866 // Only do this if the casts both really cause code to be generated.
3867 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3869 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3871 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3872 Op1C->getOperand(0),
3874 InsertNewInstBefore(NewOp, I);
3875 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3880 return Changed ? &I : 0;
3883 // XorSelf - Implements: X ^ X --> 0
3886 XorSelf(Value *rhs) : RHS(rhs) {}
3887 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3888 Instruction *apply(BinaryOperator &Xor) const {
3894 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3895 bool Changed = SimplifyCommutative(I);
3896 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3898 if (isa<UndefValue>(Op1))
3899 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3901 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3902 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3903 assert(Result == &I && "AssociativeOpt didn't work?");
3904 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3907 // See if we can simplify any instructions used by the instruction whose sole
3908 // purpose is to compute bits we don't care about.
3909 uint64_t KnownZero, KnownOne;
3910 if (!isa<PackedType>(I.getType()) &&
3911 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3912 KnownZero, KnownOne))
3915 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3916 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
3917 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
3918 if (RHS == ConstantBool::getTrue() && ICI->hasOneUse())
3919 return new ICmpInst(ICI->getInversePredicate(),
3920 ICI->getOperand(0), ICI->getOperand(1));
3922 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3923 // ~(c-X) == X-c-1 == X+(-c-1)
3924 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3925 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3926 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3927 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3928 ConstantInt::get(I.getType(), 1));
3929 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3932 // ~(~X & Y) --> (X | ~Y)
3933 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3934 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3935 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3937 BinaryOperator::createNot(Op0I->getOperand(1),
3938 Op0I->getOperand(1)->getName()+".not");
3939 InsertNewInstBefore(NotY, I);
3940 return BinaryOperator::createOr(Op0NotVal, NotY);
3944 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3945 if (Op0I->getOpcode() == Instruction::Add) {
3946 // ~(X-c) --> (-c-1)-X
3947 if (RHS->isAllOnesValue()) {
3948 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3949 return BinaryOperator::createSub(
3950 ConstantExpr::getSub(NegOp0CI,
3951 ConstantInt::get(I.getType(), 1)),
3952 Op0I->getOperand(0));
3954 } else if (Op0I->getOpcode() == Instruction::Or) {
3955 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3956 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3957 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3958 // Anything in both C1 and C2 is known to be zero, remove it from
3960 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3961 NewRHS = ConstantExpr::getAnd(NewRHS,
3962 ConstantExpr::getNot(CommonBits));
3963 WorkList.push_back(Op0I);
3964 I.setOperand(0, Op0I->getOperand(0));
3965 I.setOperand(1, NewRHS);
3971 // Try to fold constant and into select arguments.
3972 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3973 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3975 if (isa<PHINode>(Op0))
3976 if (Instruction *NV = FoldOpIntoPhi(I))
3980 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3982 return ReplaceInstUsesWith(I,
3983 ConstantIntegral::getAllOnesValue(I.getType()));
3985 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3987 return ReplaceInstUsesWith(I,
3988 ConstantIntegral::getAllOnesValue(I.getType()));
3990 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3991 if (Op1I->getOpcode() == Instruction::Or) {
3992 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3993 Op1I->swapOperands();
3995 std::swap(Op0, Op1);
3996 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3997 I.swapOperands(); // Simplified below.
3998 std::swap(Op0, Op1);
4000 } else if (Op1I->getOpcode() == Instruction::Xor) {
4001 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
4002 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
4003 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
4004 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
4005 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
4006 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
4007 Op1I->swapOperands();
4008 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
4009 I.swapOperands(); // Simplified below.
4010 std::swap(Op0, Op1);
4014 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
4015 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
4016 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
4017 Op0I->swapOperands();
4018 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
4019 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
4020 InsertNewInstBefore(NotB, I);
4021 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
4023 } else if (Op0I->getOpcode() == Instruction::Xor) {
4024 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
4025 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
4026 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
4027 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
4028 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
4029 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
4030 Op0I->swapOperands();
4031 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
4032 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4033 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
4034 InsertNewInstBefore(N, I);
4035 return BinaryOperator::createAnd(N, Op1);
4039 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4040 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4041 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4044 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4045 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4046 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4047 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4048 const Type *SrcTy = Op0C->getOperand(0)->getType();
4049 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
4050 // Only do this if the casts both really cause code to be generated.
4051 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4053 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4055 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4056 Op1C->getOperand(0),
4058 InsertNewInstBefore(NewOp, I);
4059 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4063 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4064 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
4065 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
4066 if (SI0->getOpcode() == SI1->getOpcode() &&
4067 SI0->getOperand(1) == SI1->getOperand(1) &&
4068 (SI0->hasOneUse() || SI1->hasOneUse())) {
4069 Instruction *NewOp =
4070 InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
4072 SI0->getName()), I);
4073 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
4077 return Changed ? &I : 0;
4080 static bool isPositive(ConstantInt *C) {
4081 return C->getSExtValue() >= 0;
4084 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4085 /// overflowed for this type.
4086 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4088 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
4090 if (In1->getType()->isUnsigned())
4091 return cast<ConstantInt>(Result)->getZExtValue() <
4092 cast<ConstantInt>(In1)->getZExtValue();
4093 if (isPositive(In1) != isPositive(In2))
4095 if (isPositive(In1))
4096 return cast<ConstantInt>(Result)->getSExtValue() <
4097 cast<ConstantInt>(In1)->getSExtValue();
4098 return cast<ConstantInt>(Result)->getSExtValue() >
4099 cast<ConstantInt>(In1)->getSExtValue();
4102 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4103 /// code necessary to compute the offset from the base pointer (without adding
4104 /// in the base pointer). Return the result as a signed integer of intptr size.
4105 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4106 TargetData &TD = IC.getTargetData();
4107 gep_type_iterator GTI = gep_type_begin(GEP);
4108 const Type *IntPtrTy = TD.getIntPtrType();
4109 Value *Result = Constant::getNullValue(IntPtrTy);
4111 // Build a mask for high order bits.
4112 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
4114 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4115 Value *Op = GEP->getOperand(i);
4116 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4117 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4118 if (Constant *OpC = dyn_cast<Constant>(Op)) {
4119 if (!OpC->isNullValue()) {
4120 OpC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4121 Scale = ConstantExpr::getMul(OpC, Scale);
4122 if (Constant *RC = dyn_cast<Constant>(Result))
4123 Result = ConstantExpr::getAdd(RC, Scale);
4125 // Emit an add instruction.
4126 Result = IC.InsertNewInstBefore(
4127 BinaryOperator::createAdd(Result, Scale,
4128 GEP->getName()+".offs"), I);
4132 // Convert to correct type.
4133 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, IntPtrTy,
4134 Op->getName()+".c"), I);
4136 // We'll let instcombine(mul) convert this to a shl if possible.
4137 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4138 GEP->getName()+".idx"), I);
4140 // Emit an add instruction.
4141 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4142 GEP->getName()+".offs"), I);
4148 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4149 /// else. At this point we know that the GEP is on the LHS of the comparison.
4150 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4151 ICmpInst::Predicate Cond,
4153 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4155 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4156 if (isa<PointerType>(CI->getOperand(0)->getType()))
4157 RHS = CI->getOperand(0);
4159 Value *PtrBase = GEPLHS->getOperand(0);
4160 if (PtrBase == RHS) {
4161 // As an optimization, we don't actually have to compute the actual value of
4162 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4163 // each index is zero or not.
4164 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4165 Instruction *InVal = 0;
4166 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4167 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4169 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4170 if (isa<UndefValue>(C)) // undef index -> undef.
4171 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4172 if (C->isNullValue())
4174 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4175 EmitIt = false; // This is indexing into a zero sized array?
4176 } else if (isa<ConstantInt>(C))
4177 return ReplaceInstUsesWith(I, // No comparison is needed here.
4178 ConstantBool::get(Cond == ICmpInst::ICMP_NE));
4183 new ICmpInst(Cond, GEPLHS->getOperand(i),
4184 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4188 InVal = InsertNewInstBefore(InVal, I);
4189 InsertNewInstBefore(Comp, I);
4190 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4191 InVal = BinaryOperator::createOr(InVal, Comp);
4192 else // True if all are equal
4193 InVal = BinaryOperator::createAnd(InVal, Comp);
4201 // No comparison is needed here, all indexes = 0
4202 ReplaceInstUsesWith(I, ConstantBool::get(Cond == ICmpInst::ICMP_EQ));
4205 // Only lower this if the icmp is the only user of the GEP or if we expect
4206 // the result to fold to a constant!
4207 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4208 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4209 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4210 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4211 Constant::getNullValue(Offset->getType()));
4213 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4214 // If the base pointers are different, but the indices are the same, just
4215 // compare the base pointer.
4216 if (PtrBase != GEPRHS->getOperand(0)) {
4217 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4218 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4219 GEPRHS->getOperand(0)->getType();
4221 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4222 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4223 IndicesTheSame = false;
4227 // If all indices are the same, just compare the base pointers.
4229 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4230 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4232 // Otherwise, the base pointers are different and the indices are
4233 // different, bail out.
4237 // If one of the GEPs has all zero indices, recurse.
4238 bool AllZeros = true;
4239 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4240 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4241 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4246 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4247 ICmpInst::getSwappedPredicate(Cond), I);
4249 // If the other GEP has all zero indices, recurse.
4251 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4252 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4253 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4258 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4260 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4261 // If the GEPs only differ by one index, compare it.
4262 unsigned NumDifferences = 0; // Keep track of # differences.
4263 unsigned DiffOperand = 0; // The operand that differs.
4264 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4265 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4266 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4267 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4268 // Irreconcilable differences.
4272 if (NumDifferences++) break;
4277 if (NumDifferences == 0) // SAME GEP?
4278 return ReplaceInstUsesWith(I, // No comparison is needed here.
4279 ConstantBool::get(Cond == ICmpInst::ICMP_EQ));
4280 else if (NumDifferences == 1) {
4281 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4282 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4283 if (LHSV->getType() != RHSV->getType())
4284 // Doesn't matter which one we bitconvert here.
4285 LHSV = InsertCastBefore(Instruction::BitCast, LHSV, RHSV->getType(),
4287 // Make sure we do a signed comparison here.
4288 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4292 // Only lower this if the icmp is the only user of the GEP or if we expect
4293 // the result to fold to a constant!
4294 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4295 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4296 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4297 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4298 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4299 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4305 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4306 bool Changed = SimplifyCompare(I);
4307 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4311 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
4313 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4314 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
4316 // Handle fcmp with constant RHS
4317 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4318 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4319 switch (LHSI->getOpcode()) {
4320 case Instruction::PHI:
4321 if (Instruction *NV = FoldOpIntoPhi(I))
4324 case Instruction::Select:
4325 // If either operand of the select is a constant, we can fold the
4326 // comparison into the select arms, which will cause one to be
4327 // constant folded and the select turned into a bitwise or.
4328 Value *Op1 = 0, *Op2 = 0;
4329 if (LHSI->hasOneUse()) {
4330 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4331 // Fold the known value into the constant operand.
4332 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4333 // Insert a new FCmp of the other select operand.
4334 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4335 LHSI->getOperand(2), RHSC,
4337 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4338 // Fold the known value into the constant operand.
4339 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4340 // Insert a new FCmp of the other select operand.
4341 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4342 LHSI->getOperand(1), RHSC,
4348 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4353 return Changed ? &I : 0;
4356 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4357 bool Changed = SimplifyCompare(I);
4358 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4359 const Type *Ty = Op0->getType();
4363 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
4365 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4366 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
4368 // icmp of GlobalValues can never equal each other as long as they aren't
4369 // external weak linkage type.
4370 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4371 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4372 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4373 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
4375 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4376 // addresses never equal each other! We already know that Op0 != Op1.
4377 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4378 isa<ConstantPointerNull>(Op0)) &&
4379 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4380 isa<ConstantPointerNull>(Op1)))
4381 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
4383 // icmp's with boolean values can always be turned into bitwise operations
4384 if (Ty == Type::BoolTy) {
4385 switch (I.getPredicate()) {
4386 default: assert(0 && "Invalid icmp instruction!");
4387 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4388 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4389 InsertNewInstBefore(Xor, I);
4390 return BinaryOperator::createNot(Xor);
4392 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4393 return BinaryOperator::createXor(Op0, Op1);
4395 case ICmpInst::ICMP_UGT:
4396 case ICmpInst::ICMP_SGT:
4397 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4399 case ICmpInst::ICMP_ULT:
4400 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4401 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4402 InsertNewInstBefore(Not, I);
4403 return BinaryOperator::createAnd(Not, Op1);
4405 case ICmpInst::ICMP_UGE:
4406 case ICmpInst::ICMP_SGE:
4407 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4409 case ICmpInst::ICMP_ULE:
4410 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4411 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4412 InsertNewInstBefore(Not, I);
4413 return BinaryOperator::createOr(Not, Op1);
4418 // See if we are doing a comparison between a constant and an instruction that
4419 // can be folded into the comparison.
4420 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4421 switch (I.getPredicate()) {
4423 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4424 if (CI->isMinValue(false))
4425 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4426 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4427 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4428 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4429 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4432 case ICmpInst::ICMP_SLT:
4433 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4434 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4435 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4436 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4437 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4438 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4441 case ICmpInst::ICMP_UGT:
4442 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4443 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4444 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4445 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4446 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4447 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4450 case ICmpInst::ICMP_SGT:
4451 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4452 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4453 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4454 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4455 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4456 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4459 case ICmpInst::ICMP_ULE:
4460 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4461 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4462 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4463 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4464 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4465 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4468 case ICmpInst::ICMP_SLE:
4469 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4470 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4471 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4472 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4473 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4474 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4477 case ICmpInst::ICMP_UGE:
4478 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4479 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4480 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4481 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4482 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4483 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4486 case ICmpInst::ICMP_SGE:
4487 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4488 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4489 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4490 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4491 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4492 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4496 // If we still have a icmp le or icmp ge instruction, turn it into the
4497 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4498 // already been handled above, this requires little checking.
4500 if (I.getPredicate() == ICmpInst::ICMP_ULE)
4501 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4502 if (I.getPredicate() == ICmpInst::ICMP_SLE)
4503 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4504 if (I.getPredicate() == ICmpInst::ICMP_UGE)
4505 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4506 if (I.getPredicate() == ICmpInst::ICMP_SGE)
4507 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4509 // See if we can fold the comparison based on bits known to be zero or one
4511 uint64_t KnownZero, KnownOne;
4512 if (SimplifyDemandedBits(Op0, Ty->getIntegralTypeMask(),
4513 KnownZero, KnownOne, 0))
4516 // Given the known and unknown bits, compute a range that the LHS could be
4518 if (KnownOne | KnownZero) {
4519 // Compute the Min, Max and RHS values based on the known bits. For the
4520 // EQ and NE we use unsigned values.
4521 uint64_t UMin, UMax, URHSVal;
4522 int64_t SMin, SMax, SRHSVal;
4523 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4524 SRHSVal = CI->getSExtValue();
4525 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, SMin,
4528 URHSVal = CI->getZExtValue();
4529 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, UMin,
4532 switch (I.getPredicate()) { // LE/GE have been folded already.
4533 default: assert(0 && "Unknown icmp opcode!");
4534 case ICmpInst::ICMP_EQ:
4535 if (UMax < URHSVal || UMin > URHSVal)
4536 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4538 case ICmpInst::ICMP_NE:
4539 if (UMax < URHSVal || UMin > URHSVal)
4540 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4542 case ICmpInst::ICMP_ULT:
4544 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4546 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4548 case ICmpInst::ICMP_UGT:
4550 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4552 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4554 case ICmpInst::ICMP_SLT:
4556 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4558 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4560 case ICmpInst::ICMP_SGT:
4562 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4564 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4569 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4570 // instruction, see if that instruction also has constants so that the
4571 // instruction can be folded into the icmp
4572 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4573 switch (LHSI->getOpcode()) {
4574 case Instruction::And:
4575 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4576 LHSI->getOperand(0)->hasOneUse()) {
4577 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4579 // If the LHS is an AND of a truncating cast, we can widen the
4580 // and/compare to be the input width without changing the value
4581 // produced, eliminating a cast.
4582 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4583 // We can do this transformation if either the AND constant does not
4584 // have its sign bit set or if it is an equality comparison.
4585 // Extending a relational comparison when we're checking the sign
4586 // bit would not work.
4587 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4589 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4590 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4591 ConstantInt *NewCST;
4593 if (Cast->getOperand(0)->getType()->isSigned()) {
4594 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4595 AndCST->getZExtValue());
4596 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4597 CI->getZExtValue());
4599 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4600 AndCST->getZExtValue());
4601 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4602 CI->getZExtValue());
4604 Instruction *NewAnd =
4605 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4607 InsertNewInstBefore(NewAnd, I);
4608 return new ICmpInst(I.getPredicate(), NewAnd, NewCI);
4612 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4613 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4614 // happens a LOT in code produced by the C front-end, for bitfield
4616 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
4618 // Check to see if there is a noop-cast between the shift and the and.
4620 if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
4621 if (CI->getOpcode() == Instruction::BitCast)
4622 Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
4626 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4627 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4628 const Type *AndTy = AndCST->getType(); // Type of the and.
4630 // We can fold this as long as we can't shift unknown bits
4631 // into the mask. This can only happen with signed shift
4632 // rights, as they sign-extend.
4634 bool CanFold = Shift->isLogicalShift();
4636 // To test for the bad case of the signed shr, see if any
4637 // of the bits shifted in could be tested after the mask.
4638 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4639 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4641 Constant *OShAmt = ConstantInt::get(Type::UByteTy, ShAmtVal);
4643 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4645 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4651 if (Shift->getOpcode() == Instruction::Shl)
4652 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4654 NewCst = ConstantExpr::getShl(CI, ShAmt);
4656 // Check to see if we are shifting out any of the bits being
4658 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4659 // If we shifted bits out, the fold is not going to work out.
4660 // As a special case, check to see if this means that the
4661 // result is always true or false now.
4662 if (I.getPredicate() == ICmpInst::ICMP_EQ)
4663 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4664 if (I.getPredicate() == ICmpInst::ICMP_NE)
4665 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4667 I.setOperand(1, NewCst);
4668 Constant *NewAndCST;
4669 if (Shift->getOpcode() == Instruction::Shl)
4670 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4672 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4673 LHSI->setOperand(1, NewAndCST);
4675 LHSI->setOperand(0, Shift->getOperand(0));
4677 Value *NewCast = InsertCastBefore(Instruction::BitCast,
4678 Shift->getOperand(0), AndTy,
4680 LHSI->setOperand(0, NewCast);
4682 WorkList.push_back(Shift); // Shift is dead.
4683 AddUsesToWorkList(I);
4689 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4690 // preferable because it allows the C<<Y expression to be hoisted out
4691 // of a loop if Y is invariant and X is not.
4692 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4693 I.isEquality() && !Shift->isArithmeticShift() &&
4694 isa<Instruction>(Shift->getOperand(0))) {
4697 if (Shift->getOpcode() == Instruction::LShr) {
4698 NS = new ShiftInst(Instruction::Shl, AndCST, Shift->getOperand(1),
4701 // Insert a logical shift.
4702 NS = new ShiftInst(Instruction::LShr, AndCST,
4703 Shift->getOperand(1), "tmp");
4705 InsertNewInstBefore(cast<Instruction>(NS), I);
4707 // If C's sign doesn't agree with the and, insert a cast now.
4708 if (NS->getType() != LHSI->getType())
4709 NS = InsertCastBefore(Instruction::BitCast, NS, LHSI->getType(),
4712 Value *ShiftOp = Shift->getOperand(0);
4713 if (ShiftOp->getType() != LHSI->getType())
4714 ShiftOp = InsertCastBefore(Instruction::BitCast, ShiftOp,
4715 LHSI->getType(), I);
4717 // Compute X & (C << Y).
4718 Instruction *NewAnd =
4719 BinaryOperator::createAnd(ShiftOp, NS, LHSI->getName());
4720 InsertNewInstBefore(NewAnd, I);
4722 I.setOperand(0, NewAnd);
4728 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
4729 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4730 if (I.isEquality()) {
4731 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4733 // Check that the shift amount is in range. If not, don't perform
4734 // undefined shifts. When the shift is visited it will be
4736 if (ShAmt->getZExtValue() >= TypeBits)
4739 // If we are comparing against bits always shifted out, the
4740 // comparison cannot succeed.
4742 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4743 if (Comp != CI) {// Comparing against a bit that we know is zero.
4744 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4745 Constant *Cst = ConstantBool::get(IsICMP_NE);
4746 return ReplaceInstUsesWith(I, Cst);
4749 if (LHSI->hasOneUse()) {
4750 // Otherwise strength reduce the shift into an and.
4751 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4752 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4755 if (CI->getType()->isUnsigned()) {
4756 Mask = ConstantInt::get(CI->getType(), Val);
4757 } else if (ShAmtVal != 0) {
4758 Mask = ConstantInt::get(CI->getType(), Val);
4760 Mask = ConstantInt::getAllOnesValue(CI->getType());
4764 BinaryOperator::createAnd(LHSI->getOperand(0),
4765 Mask, LHSI->getName()+".mask");
4766 Value *And = InsertNewInstBefore(AndI, I);
4767 return new ICmpInst(I.getPredicate(), And,
4768 ConstantExpr::getLShr(CI, ShAmt));
4774 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
4775 case Instruction::AShr:
4776 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4777 if (I.isEquality()) {
4778 // Check that the shift amount is in range. If not, don't perform
4779 // undefined shifts. When the shift is visited it will be
4781 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4782 if (ShAmt->getZExtValue() >= TypeBits)
4785 // If we are comparing against bits always shifted out, the
4786 // comparison cannot succeed.
4788 if (CI->getType()->isUnsigned())
4789 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4792 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4795 if (Comp != CI) {// Comparing against a bit that we know is zero.
4796 bool IsICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4797 Constant *Cst = ConstantBool::get(IsICMP_NE);
4798 return ReplaceInstUsesWith(I, Cst);
4801 if (LHSI->hasOneUse() || CI->isNullValue()) {
4802 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4804 // Otherwise strength reduce the shift into an and.
4805 uint64_t Val = ~0ULL; // All ones.
4806 Val <<= ShAmtVal; // Shift over to the right spot.
4809 if (CI->getType()->isUnsigned()) {
4810 Val &= ~0ULL >> (64-TypeBits);
4811 Mask = ConstantInt::get(CI->getType(), Val);
4813 Mask = ConstantInt::get(CI->getType(), Val);
4817 BinaryOperator::createAnd(LHSI->getOperand(0),
4818 Mask, LHSI->getName()+".mask");
4819 Value *And = InsertNewInstBefore(AndI, I);
4820 return new ICmpInst(I.getPredicate(), And,
4821 ConstantExpr::getShl(CI, ShAmt));
4827 case Instruction::SDiv:
4828 case Instruction::UDiv:
4829 // Fold: icmp pred ([us]div X, C1), C2 -> range test
4830 // Fold this div into the comparison, producing a range check.
4831 // Determine, based on the divide type, what the range is being
4832 // checked. If there is an overflow on the low or high side, remember
4833 // it, otherwise compute the range [low, hi) bounding the new value.
4834 // See: InsertRangeTest above for the kinds of replacements possible.
4835 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4836 // FIXME: If the operand types don't match the type of the divide
4837 // then don't attempt this transform. The code below doesn't have the
4838 // logic to deal with a signed divide and an unsigned compare (and
4839 // vice versa). This is because (x /s C1) <s C2 produces different
4840 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4841 // (x /u C1) <u C2. Simply casting the operands and result won't
4842 // work. :( The if statement below tests that condition and bails
4844 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
4845 if (!I.isEquality() && DivIsSigned != I.isSignedPredicate())
4848 // Initialize the variables that will indicate the nature of the
4850 bool LoOverflow = false, HiOverflow = false;
4851 ConstantInt *LoBound = 0, *HiBound = 0;
4853 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4854 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4855 // C2 (CI). By solving for X we can turn this into a range check
4856 // instead of computing a divide.
4858 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4860 // Determine if the product overflows by seeing if the product is
4861 // not equal to the divide. Make sure we do the same kind of divide
4862 // as in the LHS instruction that we're folding.
4863 bool ProdOV = !DivRHS->isNullValue() &&
4864 (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
4865 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4867 // Get the ICmp opcode
4868 ICmpInst::Predicate predicate = I.getPredicate();
4870 if (DivRHS->isNullValue()) {
4871 // Don't hack on divide by zeros!
4872 } else if (!DivIsSigned) { // udiv
4874 LoOverflow = ProdOV;
4875 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4876 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4877 if (CI->isNullValue()) { // (X / pos) op 0
4879 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4881 } else if (isPositive(CI)) { // (X / pos) op pos
4883 LoOverflow = ProdOV;
4884 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4885 } else { // (X / pos) op neg
4886 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4887 LoOverflow = AddWithOverflow(LoBound, Prod,
4888 cast<ConstantInt>(DivRHSH));
4890 HiOverflow = ProdOV;
4892 } else { // Divisor is < 0.
4893 if (CI->isNullValue()) { // (X / neg) op 0
4894 LoBound = AddOne(DivRHS);
4895 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4896 if (HiBound == DivRHS)
4897 LoBound = 0; // - INTMIN = INTMIN
4898 } else if (isPositive(CI)) { // (X / neg) op pos
4899 HiOverflow = LoOverflow = ProdOV;
4901 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4902 HiBound = AddOne(Prod);
4903 } else { // (X / neg) op neg
4905 LoOverflow = HiOverflow = ProdOV;
4906 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4909 // Dividing by a negate swaps the condition.
4910 predicate = ICmpInst::getSwappedPredicate(predicate);
4914 Value *X = LHSI->getOperand(0);
4915 switch (predicate) {
4916 default: assert(0 && "Unhandled icmp opcode!");
4917 case ICmpInst::ICMP_EQ:
4918 if (LoOverflow && HiOverflow)
4919 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4920 else if (HiOverflow)
4921 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4922 ICmpInst::ICMP_UGE, X, LoBound);
4923 else if (LoOverflow)
4924 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4925 ICmpInst::ICMP_ULT, X, HiBound);
4927 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4929 case ICmpInst::ICMP_NE:
4930 if (LoOverflow && HiOverflow)
4931 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4932 else if (HiOverflow)
4933 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
4934 ICmpInst::ICMP_ULT, X, LoBound);
4935 else if (LoOverflow)
4936 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
4937 ICmpInst::ICMP_UGE, X, HiBound);
4939 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
4941 case ICmpInst::ICMP_ULT:
4942 case ICmpInst::ICMP_SLT:
4944 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4945 return new ICmpInst(predicate, X, LoBound);
4946 case ICmpInst::ICMP_UGT:
4947 case ICmpInst::ICMP_SGT:
4949 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4950 if (predicate == ICmpInst::ICMP_UGT)
4951 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
4953 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
4960 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
4961 if (I.isEquality()) {
4962 bool isICMP_NE = I.getPredicate() == ICmpInst::ICMP_NE;
4964 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4965 // the second operand is a constant, simplify a bit.
4966 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4967 switch (BO->getOpcode()) {
4968 case Instruction::SRem:
4969 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4970 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4972 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4973 if (V > 1 && isPowerOf2_64(V)) {
4974 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
4975 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
4976 return new ICmpInst(I.getPredicate(), NewRem,
4977 Constant::getNullValue(BO->getType()));
4981 case Instruction::Add:
4982 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4983 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4984 if (BO->hasOneUse())
4985 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
4986 ConstantExpr::getSub(CI, BOp1C));
4987 } else if (CI->isNullValue()) {
4988 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4989 // efficiently invertible, or if the add has just this one use.
4990 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4992 if (Value *NegVal = dyn_castNegVal(BOp1))
4993 return new ICmpInst(I.getPredicate(), BOp0, NegVal);
4994 else if (Value *NegVal = dyn_castNegVal(BOp0))
4995 return new ICmpInst(I.getPredicate(), NegVal, BOp1);
4996 else if (BO->hasOneUse()) {
4997 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
4999 InsertNewInstBefore(Neg, I);
5000 return new ICmpInst(I.getPredicate(), BOp0, Neg);
5004 case Instruction::Xor:
5005 // For the xor case, we can xor two constants together, eliminating
5006 // the explicit xor.
5007 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5008 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5009 ConstantExpr::getXor(CI, BOC));
5012 case Instruction::Sub:
5013 // Replace (([sub|xor] A, B) != 0) with (A != B)
5014 if (CI->isNullValue())
5015 return new ICmpInst(I.getPredicate(), BO->getOperand(0),
5019 case Instruction::Or:
5020 // If bits are being or'd in that are not present in the constant we
5021 // are comparing against, then the comparison could never succeed!
5022 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5023 Constant *NotCI = ConstantExpr::getNot(CI);
5024 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5025 return ReplaceInstUsesWith(I, ConstantBool::get(isICMP_NE));
5029 case Instruction::And:
5030 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5031 // If bits are being compared against that are and'd out, then the
5032 // comparison can never succeed!
5033 if (!ConstantExpr::getAnd(CI,
5034 ConstantExpr::getNot(BOC))->isNullValue())
5035 return ReplaceInstUsesWith(I, ConstantBool::get(isICMP_NE));
5037 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5038 if (CI == BOC && isOneBitSet(CI))
5039 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5040 ICmpInst::ICMP_NE, Op0,
5041 Constant::getNullValue(CI->getType()));
5043 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5044 if (isSignBit(BOC)) {
5045 Value *X = BO->getOperand(0);
5046 // If 'X' is not signed, insert a cast now...
5047 if (!BOC->getType()->isSigned()) {
5048 const Type *DestTy = BOC->getType()->getSignedVersion();
5049 X = InsertCastBefore(Instruction::BitCast, X, DestTy, I);
5051 Constant *Zero = Constant::getNullValue(X->getType());
5052 ICmpInst::Predicate pred = isICMP_NE ?
5053 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5054 return new ICmpInst(pred, X, Zero);
5057 // ((X & ~7) == 0) --> X < 8
5058 if (CI->isNullValue() && isHighOnes(BOC)) {
5059 Value *X = BO->getOperand(0);
5060 Constant *NegX = ConstantExpr::getNeg(BOC);
5061 ICmpInst::Predicate pred = isICMP_NE ?
5062 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5063 return new ICmpInst(pred, X, NegX);
5069 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
5070 // Handle set{eq|ne} <intrinsic>, intcst.
5071 switch (II->getIntrinsicID()) {
5073 case Intrinsic::bswap_i16:
5074 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5075 WorkList.push_back(II); // Dead?
5076 I.setOperand(0, II->getOperand(1));
5077 I.setOperand(1, ConstantInt::get(Type::UShortTy,
5078 ByteSwap_16(CI->getZExtValue())));
5080 case Intrinsic::bswap_i32:
5081 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5082 WorkList.push_back(II); // Dead?
5083 I.setOperand(0, II->getOperand(1));
5084 I.setOperand(1, ConstantInt::get(Type::UIntTy,
5085 ByteSwap_32(CI->getZExtValue())));
5087 case Intrinsic::bswap_i64:
5088 // icmp eq (bswap(x)), c -> icmp eq (x,bswap(c))
5089 WorkList.push_back(II); // Dead?
5090 I.setOperand(0, II->getOperand(1));
5091 I.setOperand(1, ConstantInt::get(Type::ULongTy,
5092 ByteSwap_64(CI->getZExtValue())));
5096 } else { // Not a ICMP_EQ/ICMP_NE
5097 // If the LHS is a cast from an integral value of the same size, then
5098 // since we know the RHS is a constant, try to simlify.
5099 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
5100 Value *CastOp = Cast->getOperand(0);
5101 const Type *SrcTy = CastOp->getType();
5102 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
5103 if (SrcTy->isInteger() &&
5104 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5105 // If this is an unsigned comparison, try to make the comparison use
5106 // smaller constant values.
5107 switch (I.getPredicate()) {
5109 case ICmpInst::ICMP_ULT: { // X u< 128 => X s> -1
5110 ConstantInt *CUI = cast<ConstantInt>(CI);
5111 if (CUI->getZExtValue() == 1ULL << (SrcTySize-1))
5112 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5113 ConstantInt::get(SrcTy, -1));
5116 case ICmpInst::ICMP_UGT: { // X u> 127 => X s< 0
5117 ConstantInt *CUI = cast<ConstantInt>(CI);
5118 if (CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
5119 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5120 Constant::getNullValue(SrcTy));
5130 // Handle icmp with constant RHS
5131 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5132 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5133 switch (LHSI->getOpcode()) {
5134 case Instruction::GetElementPtr:
5135 if (RHSC->isNullValue()) {
5136 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5137 bool isAllZeros = true;
5138 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5139 if (!isa<Constant>(LHSI->getOperand(i)) ||
5140 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5145 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5146 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5150 case Instruction::PHI:
5151 if (Instruction *NV = FoldOpIntoPhi(I))
5154 case Instruction::Select:
5155 // If either operand of the select is a constant, we can fold the
5156 // comparison into the select arms, which will cause one to be
5157 // constant folded and the select turned into a bitwise or.
5158 Value *Op1 = 0, *Op2 = 0;
5159 if (LHSI->hasOneUse()) {
5160 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5161 // Fold the known value into the constant operand.
5162 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5163 // Insert a new ICmp of the other select operand.
5164 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5165 LHSI->getOperand(2), RHSC,
5167 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5168 // Fold the known value into the constant operand.
5169 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5170 // Insert a new ICmp of the other select operand.
5171 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5172 LHSI->getOperand(1), RHSC,
5178 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5183 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5184 if (User *GEP = dyn_castGetElementPtr(Op0))
5185 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5187 if (User *GEP = dyn_castGetElementPtr(Op1))
5188 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5189 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5192 // Test to see if the operands of the icmp are casted versions of other
5193 // values. If the cast can be stripped off both arguments, we do so now.
5194 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
5195 Value *CastOp0 = CI->getOperand(0);
5196 if (CI->isLosslessCast() && I.isEquality() &&
5197 (isa<Constant>(Op1) || isa<CastInst>(Op1))) {
5198 // We keep moving the cast from the left operand over to the right
5199 // operand, where it can often be eliminated completely.
5202 // If operand #1 is a cast instruction, see if we can eliminate it as
5204 if (CastInst *CI2 = dyn_cast<CastInst>(Op1)) {
5205 Value *CI2Op0 = CI2->getOperand(0);
5206 if (CI2Op0->getType()->canLosslesslyBitCastTo(Op0->getType()))
5210 // If Op1 is a constant, we can fold the cast into the constant.
5211 if (Op1->getType() != Op0->getType())
5212 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5213 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5215 // Otherwise, cast the RHS right before the icmp
5216 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5218 return new ICmpInst(I.getPredicate(), Op0, Op1);
5221 // Handle the special case of: icmp (cast bool to X), <cst>
5222 // This comes up when you have code like
5225 // For generality, we handle any zero-extension of any operand comparison
5226 // with a constant or another cast from the same type.
5227 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5228 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5232 if (I.isEquality()) {
5234 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
5235 (A == Op1 || B == Op1)) {
5236 // (A^B) == A -> B == 0
5237 Value *OtherVal = A == Op1 ? B : A;
5238 return new ICmpInst(I.getPredicate(), OtherVal,
5239 Constant::getNullValue(A->getType()));
5240 } else if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5241 (A == Op0 || B == Op0)) {
5242 // A == (A^B) -> B == 0
5243 Value *OtherVal = A == Op0 ? B : A;
5244 return new ICmpInst(I.getPredicate(), OtherVal,
5245 Constant::getNullValue(A->getType()));
5246 } else if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5247 // (A-B) == A -> B == 0
5248 return new ICmpInst(I.getPredicate(), B,
5249 Constant::getNullValue(B->getType()));
5250 } else if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5251 // A == (A-B) -> B == 0
5252 return new ICmpInst(I.getPredicate(), B,
5253 Constant::getNullValue(B->getType()));
5257 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5258 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5259 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5260 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5261 Value *X = 0, *Y = 0, *Z = 0;
5264 X = B; Y = D; Z = A;
5265 } else if (A == D) {
5266 X = B; Y = C; Z = A;
5267 } else if (B == C) {
5268 X = A; Y = D; Z = B;
5269 } else if (B == D) {
5270 X = A; Y = C; Z = B;
5273 if (X) { // Build (X^Y) & Z
5274 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5275 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5276 I.setOperand(0, Op1);
5277 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5282 return Changed ? &I : 0;
5285 // visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5286 // We only handle extending casts so far.
5288 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5289 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5290 Value *LHSCIOp = LHSCI->getOperand(0);
5291 const Type *SrcTy = LHSCIOp->getType();
5292 const Type *DestTy = LHSCI->getType();
5295 // We only handle extension cast instructions, so far. Enforce this.
5296 if (LHSCI->getOpcode() != Instruction::ZExt &&
5297 LHSCI->getOpcode() != Instruction::SExt)
5300 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5301 bool isSignedCmp = ICI.isSignedPredicate();
5303 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5304 // Not an extension from the same type?
5305 RHSCIOp = CI->getOperand(0);
5306 if (RHSCIOp->getType() != LHSCIOp->getType())
5309 // Okay, just insert a compare of the reduced operands now!
5310 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5313 // If we aren't dealing with a constant on the RHS, exit early
5314 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5318 // Compute the constant that would happen if we truncated to SrcTy then
5319 // reextended to DestTy.
5320 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5321 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5323 // If the re-extended constant didn't change...
5325 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5326 // For example, we might have:
5327 // %A = sext short %X to uint
5328 // %B = icmp ugt uint %A, 1330
5329 // It is incorrect to transform this into
5330 // %B = icmp ugt short %X, 1330
5331 // because %A may have negative value.
5333 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5334 // OR operation is EQ/NE.
5335 if (isSignedExt == isSignedCmp || SrcTy == Type::BoolTy || ICI.isEquality())
5336 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5341 // The re-extended constant changed so the constant cannot be represented
5342 // in the shorter type. Consequently, we cannot emit a simple comparison.
5344 // First, handle some easy cases. We know the result cannot be equal at this
5345 // point so handle the ICI.isEquality() cases
5346 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5347 return ReplaceInstUsesWith(ICI, ConstantBool::getFalse());
5348 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5349 return ReplaceInstUsesWith(ICI, ConstantBool::getTrue());
5351 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5352 // should have been folded away previously and not enter in here.
5355 // We're performing a signed comparison.
5356 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
5357 Result = ConstantBool::getFalse(); // X < (small) --> false
5359 Result = ConstantBool::getTrue(); // X < (large) --> true
5361 // We're performing an unsigned comparison.
5363 // We're performing an unsigned comp with a sign extended value.
5364 // This is true if the input is >= 0. [aka >s -1]
5365 Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
5366 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5367 NegOne, ICI.getName()), ICI);
5369 // Unsigned extend & unsigned compare -> always true.
5370 Result = ConstantBool::getTrue();
5374 // Finally, return the value computed.
5375 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5376 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5377 return ReplaceInstUsesWith(ICI, Result);
5379 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5380 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5381 "ICmp should be folded!");
5382 if (Constant *CI = dyn_cast<Constant>(Result))
5383 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5385 return BinaryOperator::createNot(Result);
5389 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
5390 assert(I.getOperand(1)->getType() == Type::UByteTy);
5391 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5393 // shl X, 0 == X and shr X, 0 == X
5394 // shl 0, X == 0 and shr 0, X == 0
5395 if (Op1 == Constant::getNullValue(Type::UByteTy) ||
5396 Op0 == Constant::getNullValue(Op0->getType()))
5397 return ReplaceInstUsesWith(I, Op0);
5399 if (isa<UndefValue>(Op0)) {
5400 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5401 return ReplaceInstUsesWith(I, Op0);
5402 else // undef << X -> 0, undef >>u X -> 0
5403 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5405 if (isa<UndefValue>(Op1)) {
5406 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5407 return ReplaceInstUsesWith(I, Op0);
5408 else // X << undef, X >>u undef -> 0
5409 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5412 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5413 if (I.getOpcode() == Instruction::AShr)
5414 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5415 if (CSI->isAllOnesValue())
5416 return ReplaceInstUsesWith(I, CSI);
5418 // Try to fold constant and into select arguments.
5419 if (isa<Constant>(Op0))
5420 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5421 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5424 // See if we can turn a signed shr into an unsigned shr.
5425 if (I.isArithmeticShift()) {
5426 if (MaskedValueIsZero(Op0,
5427 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5428 return new ShiftInst(Instruction::LShr, Op0, Op1, I.getName());
5432 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5433 if (CUI->getType()->isUnsigned())
5434 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5439 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5441 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5442 bool isSignedShift = I.getOpcode() == Instruction::AShr;
5443 bool isUnsignedShift = !isSignedShift;
5445 // See if we can simplify any instructions used by the instruction whose sole
5446 // purpose is to compute bits we don't care about.
5447 uint64_t KnownZero, KnownOne;
5448 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
5449 KnownZero, KnownOne))
5452 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5453 // of a signed value.
5455 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5456 if (Op1->getZExtValue() >= TypeBits) {
5457 if (isUnsignedShift || isLeftShift)
5458 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5460 I.setOperand(1, ConstantInt::get(Type::UByteTy, TypeBits-1));
5465 // ((X*C1) << C2) == (X * (C1 << C2))
5466 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5467 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5468 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5469 return BinaryOperator::createMul(BO->getOperand(0),
5470 ConstantExpr::getShl(BOOp, Op1));
5472 // Try to fold constant and into select arguments.
5473 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5474 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5476 if (isa<PHINode>(Op0))
5477 if (Instruction *NV = FoldOpIntoPhi(I))
5480 if (Op0->hasOneUse()) {
5481 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5482 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5485 switch (Op0BO->getOpcode()) {
5487 case Instruction::Add:
5488 case Instruction::And:
5489 case Instruction::Or:
5490 case Instruction::Xor:
5491 // These operators commute.
5492 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5493 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5494 match(Op0BO->getOperand(1),
5495 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5496 Instruction *YS = new ShiftInst(Instruction::Shl,
5497 Op0BO->getOperand(0), Op1,
5499 InsertNewInstBefore(YS, I); // (Y << C)
5501 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5502 Op0BO->getOperand(1)->getName());
5503 InsertNewInstBefore(X, I); // (X + (Y << C))
5504 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5505 C2 = ConstantExpr::getShl(C2, Op1);
5506 return BinaryOperator::createAnd(X, C2);
5509 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5510 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5511 match(Op0BO->getOperand(1),
5512 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5513 m_ConstantInt(CC))) && V2 == Op1 &&
5514 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
5515 Instruction *YS = new ShiftInst(Instruction::Shl,
5516 Op0BO->getOperand(0), Op1,
5518 InsertNewInstBefore(YS, I); // (Y << C)
5520 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5521 V1->getName()+".mask");
5522 InsertNewInstBefore(XM, I); // X & (CC << C)
5524 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5528 case Instruction::Sub:
5529 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5530 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5531 match(Op0BO->getOperand(0),
5532 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5533 Instruction *YS = new ShiftInst(Instruction::Shl,
5534 Op0BO->getOperand(1), Op1,
5536 InsertNewInstBefore(YS, I); // (Y << C)
5538 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5539 Op0BO->getOperand(0)->getName());
5540 InsertNewInstBefore(X, I); // (X + (Y << C))
5541 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5542 C2 = ConstantExpr::getShl(C2, Op1);
5543 return BinaryOperator::createAnd(X, C2);
5546 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5547 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5548 match(Op0BO->getOperand(0),
5549 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5550 m_ConstantInt(CC))) && V2 == Op1 &&
5551 cast<BinaryOperator>(Op0BO->getOperand(0))
5552 ->getOperand(0)->hasOneUse()) {
5553 Instruction *YS = new ShiftInst(Instruction::Shl,
5554 Op0BO->getOperand(1), Op1,
5556 InsertNewInstBefore(YS, I); // (Y << C)
5558 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5559 V1->getName()+".mask");
5560 InsertNewInstBefore(XM, I); // X & (CC << C)
5562 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5569 // If the operand is an bitwise operator with a constant RHS, and the
5570 // shift is the only use, we can pull it out of the shift.
5571 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5572 bool isValid = true; // Valid only for And, Or, Xor
5573 bool highBitSet = false; // Transform if high bit of constant set?
5575 switch (Op0BO->getOpcode()) {
5576 default: isValid = false; break; // Do not perform transform!
5577 case Instruction::Add:
5578 isValid = isLeftShift;
5580 case Instruction::Or:
5581 case Instruction::Xor:
5584 case Instruction::And:
5589 // If this is a signed shift right, and the high bit is modified
5590 // by the logical operation, do not perform the transformation.
5591 // The highBitSet boolean indicates the value of the high bit of
5592 // the constant which would cause it to be modified for this
5595 if (isValid && !isLeftShift && isSignedShift) {
5596 uint64_t Val = Op0C->getZExtValue();
5597 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5601 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5603 Instruction *NewShift =
5604 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
5607 InsertNewInstBefore(NewShift, I);
5609 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5616 // Find out if this is a shift of a shift by a constant.
5617 ShiftInst *ShiftOp = 0;
5618 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
5620 else if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5621 // If this is a noop-integer cast of a shift instruction, use the shift.
5622 if (isa<ShiftInst>(CI->getOperand(0))) {
5623 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
5627 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5628 // Find the operands and properties of the input shift. Note that the
5629 // signedness of the input shift may differ from the current shift if there
5630 // is a noop cast between the two.
5631 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
5632 bool isShiftOfSignedShift = ShiftOp->getOpcode() == Instruction::AShr;
5633 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
5635 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5637 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5638 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5640 // Check for (A << c1) << c2 and (A >> c1) >> c2.
5641 if (isLeftShift == isShiftOfLeftShift) {
5642 // Do not fold these shifts if the first one is signed and the second one
5643 // is unsigned and this is a right shift. Further, don't do any folding
5645 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
5648 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5649 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
5650 Amt = Op0->getType()->getPrimitiveSizeInBits();
5652 Value *Op = ShiftOp->getOperand(0);
5653 ShiftInst *ShiftResult = new ShiftInst(I.getOpcode(), Op,
5654 ConstantInt::get(Type::UByteTy, Amt));
5655 if (I.getType() == ShiftResult->getType())
5657 InsertNewInstBefore(ShiftResult, I);
5658 return CastInst::create(Instruction::BitCast, ShiftResult, I.getType());
5661 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
5662 // signed types, we can only support the (A >> c1) << c2 configuration,
5663 // because it can not turn an arbitrary bit of A into a sign bit.
5664 if (isUnsignedShift || isLeftShift) {
5665 // Calculate bitmask for what gets shifted off the edge.
5666 Constant *C = ConstantIntegral::getAllOnesValue(I.getType());
5668 C = ConstantExpr::getShl(C, ShiftAmt1C);
5670 C = ConstantExpr::getLShr(C, ShiftAmt1C);
5672 Value *Op = ShiftOp->getOperand(0);
5673 if (Op->getType() != C->getType())
5674 Op = InsertCastBefore(Instruction::BitCast, Op, I.getType(), I);
5677 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
5678 InsertNewInstBefore(Mask, I);
5680 // Figure out what flavor of shift we should use...
5681 if (ShiftAmt1 == ShiftAmt2) {
5682 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
5683 } else if (ShiftAmt1 < ShiftAmt2) {
5684 return new ShiftInst(I.getOpcode(), Mask,
5685 ConstantInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
5686 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
5687 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
5688 return new ShiftInst(Instruction::LShr, Mask,
5689 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5691 return new ShiftInst(ShiftOp->getOpcode(), Mask,
5692 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5695 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
5696 Instruction *Shift =
5697 new ShiftInst(ShiftOp->getOpcode(), Mask,
5698 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5699 InsertNewInstBefore(Shift, I);
5701 C = ConstantIntegral::getAllOnesValue(Shift->getType());
5702 C = ConstantExpr::getShl(C, Op1);
5703 return BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
5706 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
5707 // this case, C1 == C2 and C1 is 8, 16, or 32.
5708 if (ShiftAmt1 == ShiftAmt2) {
5709 const Type *SExtType = 0;
5710 switch (Op0->getType()->getPrimitiveSizeInBits() - ShiftAmt1) {
5711 case 8 : SExtType = Type::SByteTy; break;
5712 case 16: SExtType = Type::ShortTy; break;
5713 case 32: SExtType = Type::IntTy; break;
5717 Instruction *NewTrunc =
5718 new TruncInst(ShiftOp->getOperand(0), SExtType, "sext");
5719 InsertNewInstBefore(NewTrunc, I);
5720 return new SExtInst(NewTrunc, I.getType());
5729 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5730 /// expression. If so, decompose it, returning some value X, such that Val is
5733 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5735 assert(Val->getType() == Type::UIntTy && "Unexpected allocation size type!");
5736 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5737 if (CI->getType()->isUnsigned()) {
5738 Offset = CI->getZExtValue();
5740 return ConstantInt::get(Type::UIntTy, 0);
5742 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5743 if (I->getNumOperands() == 2) {
5744 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5745 if (CUI->getType()->isUnsigned()) {
5746 if (I->getOpcode() == Instruction::Shl) {
5747 // This is a value scaled by '1 << the shift amt'.
5748 Scale = 1U << CUI->getZExtValue();
5750 return I->getOperand(0);
5751 } else if (I->getOpcode() == Instruction::Mul) {
5752 // This value is scaled by 'CUI'.
5753 Scale = CUI->getZExtValue();
5755 return I->getOperand(0);
5756 } else if (I->getOpcode() == Instruction::Add) {
5757 // We have X+C. Check to see if we really have (X*C2)+C1,
5758 // where C1 is divisible by C2.
5761 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5762 Offset += CUI->getZExtValue();
5763 if (SubScale > 1 && (Offset % SubScale == 0)) {
5773 // Otherwise, we can't look past this.
5780 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5781 /// try to eliminate the cast by moving the type information into the alloc.
5782 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5783 AllocationInst &AI) {
5784 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5785 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5787 // Remove any uses of AI that are dead.
5788 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5789 std::vector<Instruction*> DeadUsers;
5790 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5791 Instruction *User = cast<Instruction>(*UI++);
5792 if (isInstructionTriviallyDead(User)) {
5793 while (UI != E && *UI == User)
5794 ++UI; // If this instruction uses AI more than once, don't break UI.
5796 // Add operands to the worklist.
5797 AddUsesToWorkList(*User);
5799 DOUT << "IC: DCE: " << *User;
5801 User->eraseFromParent();
5802 removeFromWorkList(User);
5806 // Get the type really allocated and the type casted to.
5807 const Type *AllocElTy = AI.getAllocatedType();
5808 const Type *CastElTy = PTy->getElementType();
5809 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5811 unsigned AllocElTyAlign = TD->getTypeAlignment(AllocElTy);
5812 unsigned CastElTyAlign = TD->getTypeAlignment(CastElTy);
5813 if (CastElTyAlign < AllocElTyAlign) return 0;
5815 // If the allocation has multiple uses, only promote it if we are strictly
5816 // increasing the alignment of the resultant allocation. If we keep it the
5817 // same, we open the door to infinite loops of various kinds.
5818 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5820 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5821 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5822 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5824 // See if we can satisfy the modulus by pulling a scale out of the array
5826 unsigned ArraySizeScale, ArrayOffset;
5827 Value *NumElements = // See if the array size is a decomposable linear expr.
5828 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5830 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5832 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5833 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5835 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5840 // If the allocation size is constant, form a constant mul expression
5841 Amt = ConstantInt::get(Type::UIntTy, Scale);
5842 if (isa<ConstantInt>(NumElements) && NumElements->getType()->isUnsigned())
5843 Amt = ConstantExpr::getMul(
5844 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5845 // otherwise multiply the amount and the number of elements
5846 else if (Scale != 1) {
5847 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5848 Amt = InsertNewInstBefore(Tmp, AI);
5852 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5853 Value *Off = ConstantInt::get(Type::UIntTy, Offset);
5854 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5855 Amt = InsertNewInstBefore(Tmp, AI);
5858 std::string Name = AI.getName(); AI.setName("");
5859 AllocationInst *New;
5860 if (isa<MallocInst>(AI))
5861 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
5863 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
5864 InsertNewInstBefore(New, AI);
5866 // If the allocation has multiple uses, insert a cast and change all things
5867 // that used it to use the new cast. This will also hack on CI, but it will
5869 if (!AI.hasOneUse()) {
5870 AddUsesToWorkList(AI);
5871 // New is the allocation instruction, pointer typed. AI is the original
5872 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
5873 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
5874 InsertNewInstBefore(NewCast, AI);
5875 AI.replaceAllUsesWith(NewCast);
5877 return ReplaceInstUsesWith(CI, New);
5880 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5881 /// and return it without inserting any new casts. This is used by code that
5882 /// tries to decide whether promoting or shrinking integer operations to wider
5883 /// or smaller types will allow us to eliminate a truncate or extend.
5884 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5885 int &NumCastsRemoved) {
5886 if (isa<Constant>(V)) return true;
5888 Instruction *I = dyn_cast<Instruction>(V);
5889 if (!I || !I->hasOneUse()) return false;
5891 switch (I->getOpcode()) {
5892 case Instruction::And:
5893 case Instruction::Or:
5894 case Instruction::Xor:
5895 // These operators can all arbitrarily be extended or truncated.
5896 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5897 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5898 case Instruction::AShr:
5899 case Instruction::LShr:
5900 case Instruction::Shl:
5901 // If this is just a bitcast changing the sign of the operation, we can
5902 // convert if the operand can be converted.
5903 if (V->getType()->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
5904 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
5906 case Instruction::Trunc:
5907 case Instruction::ZExt:
5908 case Instruction::SExt:
5909 case Instruction::BitCast:
5910 // If this is a cast from the destination type, we can trivially eliminate
5911 // it, and this will remove a cast overall.
5912 if (I->getOperand(0)->getType() == Ty) {
5913 // If the first operand is itself a cast, and is eliminable, do not count
5914 // this as an eliminable cast. We would prefer to eliminate those two
5916 if (isa<CastInst>(I->getOperand(0)))
5924 // TODO: Can handle more cases here.
5931 /// EvaluateInDifferentType - Given an expression that
5932 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5933 /// evaluate the expression.
5934 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
5936 if (Constant *C = dyn_cast<Constant>(V))
5937 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
5939 // Otherwise, it must be an instruction.
5940 Instruction *I = cast<Instruction>(V);
5941 Instruction *Res = 0;
5942 switch (I->getOpcode()) {
5943 case Instruction::And:
5944 case Instruction::Or:
5945 case Instruction::Xor: {
5946 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5947 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
5948 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5949 LHS, RHS, I->getName());
5952 case Instruction::AShr:
5953 case Instruction::LShr:
5954 case Instruction::Shl: {
5955 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5956 Res = new ShiftInst((Instruction::OtherOps)I->getOpcode(), LHS,
5957 I->getOperand(1), I->getName());
5960 case Instruction::Trunc:
5961 case Instruction::ZExt:
5962 case Instruction::SExt:
5963 case Instruction::BitCast:
5964 // If the source type of the cast is the type we're trying for then we can
5965 // just return the source. There's no need to insert it because its not new.
5966 if (I->getOperand(0)->getType() == Ty)
5967 return I->getOperand(0);
5969 // Some other kind of cast, which shouldn't happen, so just ..
5972 // TODO: Can handle more cases here.
5973 assert(0 && "Unreachable!");
5977 return InsertNewInstBefore(Res, *I);
5980 /// @brief Implement the transforms common to all CastInst visitors.
5981 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
5982 Value *Src = CI.getOperand(0);
5984 // Casting undef to anything results in undef so might as just replace it and
5985 // get rid of the cast.
5986 if (isa<UndefValue>(Src)) // cast undef -> undef
5987 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5989 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
5990 // eliminate it now.
5991 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5992 if (Instruction::CastOps opc =
5993 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
5994 // The first cast (CSrc) is eliminable so we need to fix up or replace
5995 // the second cast (CI). CSrc will then have a good chance of being dead.
5996 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6000 // If casting the result of a getelementptr instruction with no offset, turn
6001 // this into a cast of the original pointer!
6003 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6004 bool AllZeroOperands = true;
6005 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
6006 if (!isa<Constant>(GEP->getOperand(i)) ||
6007 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
6008 AllZeroOperands = false;
6011 if (AllZeroOperands) {
6012 // Changing the cast operand is usually not a good idea but it is safe
6013 // here because the pointer operand is being replaced with another
6014 // pointer operand so the opcode doesn't need to change.
6015 CI.setOperand(0, GEP->getOperand(0));
6020 // If we are casting a malloc or alloca to a pointer to a type of the same
6021 // size, rewrite the allocation instruction to allocate the "right" type.
6022 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
6023 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6026 // If we are casting a select then fold the cast into the select
6027 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6028 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6031 // If we are casting a PHI then fold the cast into the PHI
6032 if (isa<PHINode>(Src))
6033 if (Instruction *NV = FoldOpIntoPhi(CI))
6039 /// Only the TRUNC, ZEXT, SEXT, and BITCONVERT can have both operands as
6040 /// integers. This function implements the common transforms for all those
6042 /// @brief Implement the transforms common to CastInst with integer operands
6043 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6044 if (Instruction *Result = commonCastTransforms(CI))
6047 Value *Src = CI.getOperand(0);
6048 const Type *SrcTy = Src->getType();
6049 const Type *DestTy = CI.getType();
6050 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6051 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
6053 // See if we can simplify any instructions used by the LHS whose sole
6054 // purpose is to compute bits we don't care about.
6055 uint64_t KnownZero = 0, KnownOne = 0;
6056 if (SimplifyDemandedBits(&CI, DestTy->getIntegralTypeMask(),
6057 KnownZero, KnownOne))
6060 // If the source isn't an instruction or has more than one use then we
6061 // can't do anything more.
6062 Instruction *SrcI = dyn_cast<Instruction>(Src);
6063 if (!SrcI || !Src->hasOneUse())
6066 // Attempt to propagate the cast into the instruction.
6067 int NumCastsRemoved = 0;
6068 if (CanEvaluateInDifferentType(SrcI, DestTy, NumCastsRemoved)) {
6069 // If this cast is a truncate, evaluting in a different type always
6070 // eliminates the cast, so it is always a win. If this is a noop-cast
6071 // this just removes a noop cast which isn't pointful, but simplifies
6072 // the code. If this is a zero-extension, we need to do an AND to
6073 // maintain the clear top-part of the computation, so we require that
6074 // the input have eliminated at least one cast. If this is a sign
6075 // extension, we insert two new casts (to do the extension) so we
6076 // require that two casts have been eliminated.
6077 bool DoXForm = CI.isNoopCast(TD->getIntPtrType());
6079 switch (CI.getOpcode()) {
6080 case Instruction::Trunc:
6083 case Instruction::ZExt:
6084 DoXForm = NumCastsRemoved >= 1;
6086 case Instruction::SExt:
6087 DoXForm = NumCastsRemoved >= 2;
6089 case Instruction::BitCast:
6093 // All the others use floating point so we shouldn't actually
6094 // get here because of the check above.
6095 assert(!"Unknown cast type .. unreachable");
6101 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6102 CI.getOpcode() == Instruction::SExt);
6103 assert(Res->getType() == DestTy);
6104 switch (CI.getOpcode()) {
6105 default: assert(0 && "Unknown cast type!");
6106 case Instruction::Trunc:
6107 case Instruction::BitCast:
6108 // Just replace this cast with the result.
6109 return ReplaceInstUsesWith(CI, Res);
6110 case Instruction::ZExt: {
6111 // We need to emit an AND to clear the high bits.
6112 assert(SrcBitSize < DestBitSize && "Not a zext?");
6114 ConstantInt::get(Type::ULongTy, (1ULL << SrcBitSize)-1);
6115 if (DestBitSize < 64)
6116 C = ConstantExpr::getTrunc(C, DestTy);
6118 assert(DestBitSize == 64);
6119 C = ConstantExpr::getBitCast(C, DestTy);
6121 return BinaryOperator::createAnd(Res, C);
6123 case Instruction::SExt:
6124 // We need to emit a cast to truncate, then a cast to sext.
6125 return CastInst::create(Instruction::SExt,
6126 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6132 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6133 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6135 switch (SrcI->getOpcode()) {
6136 case Instruction::Add:
6137 case Instruction::Mul:
6138 case Instruction::And:
6139 case Instruction::Or:
6140 case Instruction::Xor:
6141 // If we are discarding information, or just changing the sign,
6143 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6144 // Don't insert two casts if they cannot be eliminated. We allow
6145 // two casts to be inserted if the sizes are the same. This could
6146 // only be converting signedness, which is a noop.
6147 if (DestBitSize == SrcBitSize ||
6148 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6149 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6150 Instruction::CastOps opcode = CI.getOpcode();
6151 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6152 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6153 return BinaryOperator::create(
6154 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6158 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6159 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6160 SrcI->getOpcode() == Instruction::Xor &&
6161 Op1 == ConstantBool::getTrue() &&
6162 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6163 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6164 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6167 case Instruction::SDiv:
6168 case Instruction::UDiv:
6169 case Instruction::SRem:
6170 case Instruction::URem:
6171 // If we are just changing the sign, rewrite.
6172 if (DestBitSize == SrcBitSize) {
6173 // Don't insert two casts if they cannot be eliminated. We allow
6174 // two casts to be inserted if the sizes are the same. This could
6175 // only be converting signedness, which is a noop.
6176 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6177 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6178 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6180 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6182 return BinaryOperator::create(
6183 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6188 case Instruction::Shl:
6189 // Allow changing the sign of the source operand. Do not allow
6190 // changing the size of the shift, UNLESS the shift amount is a
6191 // constant. We must not change variable sized shifts to a smaller
6192 // size, because it is undefined to shift more bits out than exist
6194 if (DestBitSize == SrcBitSize ||
6195 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6196 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6197 Instruction::BitCast : Instruction::Trunc);
6198 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6199 return new ShiftInst(Instruction::Shl, Op0c, Op1);
6202 case Instruction::AShr:
6203 // If this is a signed shr, and if all bits shifted in are about to be
6204 // truncated off, turn it into an unsigned shr to allow greater
6206 if (DestBitSize < SrcBitSize &&
6207 isa<ConstantInt>(Op1)) {
6208 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
6209 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6210 // Insert the new logical shift right.
6211 return new ShiftInst(Instruction::LShr, Op0, Op1);
6216 case Instruction::ICmp:
6217 // If we are just checking for a icmp eq of a single bit and casting it
6218 // to an integer, then shift the bit to the appropriate place and then
6219 // cast to integer to avoid the comparison.
6220 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
6221 uint64_t Op1CV = Op1C->getZExtValue();
6222 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
6223 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6224 // cast (X == 1) to int --> X iff X has only the low bit set.
6225 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
6226 // cast (X != 0) to int --> X iff X has only the low bit set.
6227 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
6228 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
6229 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
6230 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
6231 // If Op1C some other power of two, convert:
6232 uint64_t KnownZero, KnownOne;
6233 uint64_t TypeMask = Op1->getType()->getIntegralTypeMask();
6234 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
6236 // This only works for EQ and NE
6237 ICmpInst::Predicate pred = cast<ICmpInst>(SrcI)->getPredicate();
6238 if (pred != ICmpInst::ICMP_NE && pred != ICmpInst::ICMP_EQ)
6241 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
6242 bool isNE = pred == ICmpInst::ICMP_NE;
6243 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
6244 // (X&4) == 2 --> false
6245 // (X&4) != 2 --> true
6246 Constant *Res = ConstantBool::get(isNE);
6247 Res = ConstantExpr::getZExt(Res, CI.getType());
6248 return ReplaceInstUsesWith(CI, Res);
6251 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
6254 // Perform a logical shr by shiftamt.
6255 // Insert the shift to put the result in the low bit.
6256 In = InsertNewInstBefore(
6257 new ShiftInst(Instruction::LShr, In,
6258 ConstantInt::get(Type::UByteTy, ShiftAmt),
6259 In->getName()+".lobit"), CI);
6262 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6263 Constant *One = ConstantInt::get(In->getType(), 1);
6264 In = BinaryOperator::createXor(In, One, "tmp");
6265 InsertNewInstBefore(cast<Instruction>(In), CI);
6268 if (CI.getType() == In->getType())
6269 return ReplaceInstUsesWith(CI, In);
6271 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6280 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
6281 if (Instruction *Result = commonIntCastTransforms(CI))
6284 Value *Src = CI.getOperand(0);
6285 const Type *Ty = CI.getType();
6286 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6288 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6289 switch (SrcI->getOpcode()) {
6291 case Instruction::LShr:
6292 // We can shrink lshr to something smaller if we know the bits shifted in
6293 // are already zeros.
6294 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6295 unsigned ShAmt = ShAmtV->getZExtValue();
6297 // Get a mask for the bits shifting in.
6298 uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
6299 Value* SrcIOp0 = SrcI->getOperand(0);
6300 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6301 if (ShAmt >= DestBitWidth) // All zeros.
6302 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6304 // Okay, we can shrink this. Truncate the input, then return a new
6306 Value *V = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6307 return new ShiftInst(Instruction::LShr, V, SrcI->getOperand(1));
6309 } else { // This is a variable shr.
6311 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6312 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6313 // loop-invariant and CSE'd.
6314 if (CI.getType() == Type::BoolTy && SrcI->hasOneUse()) {
6315 Value *One = ConstantInt::get(SrcI->getType(), 1);
6317 Value *V = InsertNewInstBefore(new ShiftInst(Instruction::Shl, One,
6318 SrcI->getOperand(1),
6320 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6321 SrcI->getOperand(0),
6323 Value *Zero = Constant::getNullValue(V->getType());
6324 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6334 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6335 // If one of the common conversion will work ..
6336 if (Instruction *Result = commonIntCastTransforms(CI))
6339 Value *Src = CI.getOperand(0);
6341 // If this is a cast of a cast
6342 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6343 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6344 // types and if the sizes are just right we can convert this into a logical
6345 // 'and' which will be much cheaper than the pair of casts.
6346 if (isa<TruncInst>(CSrc)) {
6347 // Get the sizes of the types involved
6348 Value *A = CSrc->getOperand(0);
6349 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6350 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6351 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6352 // If we're actually extending zero bits and the trunc is a no-op
6353 if (MidSize < DstSize && SrcSize == DstSize) {
6354 // Replace both of the casts with an And of the type mask.
6355 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
6356 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
6358 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6359 // Unfortunately, if the type changed, we need to cast it back.
6360 if (And->getType() != CI.getType()) {
6361 And->setName(CSrc->getName()+".mask");
6362 InsertNewInstBefore(And, CI);
6363 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6373 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6374 return commonIntCastTransforms(CI);
6377 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6378 return commonCastTransforms(CI);
6381 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6382 return commonCastTransforms(CI);
6385 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6386 return commonCastTransforms(CI);
6389 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6390 return commonCastTransforms(CI);
6393 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6394 return commonCastTransforms(CI);
6397 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6398 return commonCastTransforms(CI);
6401 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6402 return commonCastTransforms(CI);
6405 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6406 return commonCastTransforms(CI);
6409 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6411 // If the operands are integer typed then apply the integer transforms,
6412 // otherwise just apply the common ones.
6413 Value *Src = CI.getOperand(0);
6414 const Type *SrcTy = Src->getType();
6415 const Type *DestTy = CI.getType();
6417 if (SrcTy->isInteger() && DestTy->isInteger()) {
6418 if (Instruction *Result = commonIntCastTransforms(CI))
6421 if (Instruction *Result = commonCastTransforms(CI))
6426 // Get rid of casts from one type to the same type. These are useless and can
6427 // be replaced by the operand.
6428 if (DestTy == Src->getType())
6429 return ReplaceInstUsesWith(CI, Src);
6431 // If the source and destination are pointers, and this cast is equivalent to
6432 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6433 // This can enhance SROA and other transforms that want type-safe pointers.
6434 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6435 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6436 const Type *DstElTy = DstPTy->getElementType();
6437 const Type *SrcElTy = SrcPTy->getElementType();
6439 Constant *ZeroUInt = Constant::getNullValue(Type::UIntTy);
6440 unsigned NumZeros = 0;
6441 while (SrcElTy != DstElTy &&
6442 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6443 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6444 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6448 // If we found a path from the src to dest, create the getelementptr now.
6449 if (SrcElTy == DstElTy) {
6450 std::vector<Value*> Idxs(NumZeros+1, ZeroUInt);
6451 return new GetElementPtrInst(Src, Idxs);
6456 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6457 if (SVI->hasOneUse()) {
6458 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6459 // a bitconvert to a vector with the same # elts.
6460 if (isa<PackedType>(DestTy) &&
6461 cast<PackedType>(DestTy)->getNumElements() ==
6462 SVI->getType()->getNumElements()) {
6464 // If either of the operands is a cast from CI.getType(), then
6465 // evaluating the shuffle in the casted destination's type will allow
6466 // us to eliminate at least one cast.
6467 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6468 Tmp->getOperand(0)->getType() == DestTy) ||
6469 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6470 Tmp->getOperand(0)->getType() == DestTy)) {
6471 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6472 SVI->getOperand(0), DestTy, &CI);
6473 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6474 SVI->getOperand(1), DestTy, &CI);
6475 // Return a new shuffle vector. Use the same element ID's, as we
6476 // know the vector types match #elts.
6477 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6485 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6487 /// %D = select %cond, %C, %A
6489 /// %C = select %cond, %B, 0
6492 /// Assuming that the specified instruction is an operand to the select, return
6493 /// a bitmask indicating which operands of this instruction are foldable if they
6494 /// equal the other incoming value of the select.
6496 static unsigned GetSelectFoldableOperands(Instruction *I) {
6497 switch (I->getOpcode()) {
6498 case Instruction::Add:
6499 case Instruction::Mul:
6500 case Instruction::And:
6501 case Instruction::Or:
6502 case Instruction::Xor:
6503 return 3; // Can fold through either operand.
6504 case Instruction::Sub: // Can only fold on the amount subtracted.
6505 case Instruction::Shl: // Can only fold on the shift amount.
6506 case Instruction::LShr:
6507 case Instruction::AShr:
6510 return 0; // Cannot fold
6514 /// GetSelectFoldableConstant - For the same transformation as the previous
6515 /// function, return the identity constant that goes into the select.
6516 static Constant *GetSelectFoldableConstant(Instruction *I) {
6517 switch (I->getOpcode()) {
6518 default: assert(0 && "This cannot happen!"); abort();
6519 case Instruction::Add:
6520 case Instruction::Sub:
6521 case Instruction::Or:
6522 case Instruction::Xor:
6523 return Constant::getNullValue(I->getType());
6524 case Instruction::Shl:
6525 case Instruction::LShr:
6526 case Instruction::AShr:
6527 return Constant::getNullValue(Type::UByteTy);
6528 case Instruction::And:
6529 return ConstantInt::getAllOnesValue(I->getType());
6530 case Instruction::Mul:
6531 return ConstantInt::get(I->getType(), 1);
6535 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6536 /// have the same opcode and only one use each. Try to simplify this.
6537 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6539 if (TI->getNumOperands() == 1) {
6540 // If this is a non-volatile load or a cast from the same type,
6543 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6546 return 0; // unknown unary op.
6549 // Fold this by inserting a select from the input values.
6550 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6551 FI->getOperand(0), SI.getName()+".v");
6552 InsertNewInstBefore(NewSI, SI);
6553 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6557 // Only handle binary, compare and shift operators here.
6558 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI) && !isa<CmpInst>(TI))
6561 // If the CmpInst predicates don't match, then the instructions aren't the
6562 // same and we can't continue.
6563 if (isa<CmpInst>(TI) && isa<CmpInst>(FI) &&
6564 (cast<CmpInst>(TI)->getPredicate() != cast<CmpInst>(FI)->getPredicate()))
6567 // Figure out if the operations have any operands in common.
6568 Value *MatchOp, *OtherOpT, *OtherOpF;
6570 if (TI->getOperand(0) == FI->getOperand(0)) {
6571 MatchOp = TI->getOperand(0);
6572 OtherOpT = TI->getOperand(1);
6573 OtherOpF = FI->getOperand(1);
6574 MatchIsOpZero = true;
6575 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6576 MatchOp = TI->getOperand(1);
6577 OtherOpT = TI->getOperand(0);
6578 OtherOpF = FI->getOperand(0);
6579 MatchIsOpZero = false;
6580 } else if (!TI->isCommutative()) {
6582 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6583 MatchOp = TI->getOperand(0);
6584 OtherOpT = TI->getOperand(1);
6585 OtherOpF = FI->getOperand(0);
6586 MatchIsOpZero = true;
6587 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6588 MatchOp = TI->getOperand(1);
6589 OtherOpT = TI->getOperand(0);
6590 OtherOpF = FI->getOperand(1);
6591 MatchIsOpZero = true;
6596 // If we reach here, they do have operations in common.
6597 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6598 OtherOpF, SI.getName()+".v");
6599 InsertNewInstBefore(NewSI, SI);
6601 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6603 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6605 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6608 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
6610 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
6614 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6615 Value *CondVal = SI.getCondition();
6616 Value *TrueVal = SI.getTrueValue();
6617 Value *FalseVal = SI.getFalseValue();
6619 // select true, X, Y -> X
6620 // select false, X, Y -> Y
6621 if (ConstantBool *C = dyn_cast<ConstantBool>(CondVal))
6622 return ReplaceInstUsesWith(SI, C->getValue() ? TrueVal : FalseVal);
6624 // select C, X, X -> X
6625 if (TrueVal == FalseVal)
6626 return ReplaceInstUsesWith(SI, TrueVal);
6628 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6629 return ReplaceInstUsesWith(SI, FalseVal);
6630 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6631 return ReplaceInstUsesWith(SI, TrueVal);
6632 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6633 if (isa<Constant>(TrueVal))
6634 return ReplaceInstUsesWith(SI, TrueVal);
6636 return ReplaceInstUsesWith(SI, FalseVal);
6639 if (SI.getType() == Type::BoolTy)
6640 if (ConstantBool *C = dyn_cast<ConstantBool>(TrueVal)) {
6641 if (C->getValue()) {
6642 // Change: A = select B, true, C --> A = or B, C
6643 return BinaryOperator::createOr(CondVal, FalseVal);
6645 // Change: A = select B, false, C --> A = and !B, C
6647 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6648 "not."+CondVal->getName()), SI);
6649 return BinaryOperator::createAnd(NotCond, FalseVal);
6651 } else if (ConstantBool *C = dyn_cast<ConstantBool>(FalseVal)) {
6652 if (C->getValue() == false) {
6653 // Change: A = select B, C, false --> A = and B, C
6654 return BinaryOperator::createAnd(CondVal, TrueVal);
6656 // Change: A = select B, C, true --> A = or !B, C
6658 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6659 "not."+CondVal->getName()), SI);
6660 return BinaryOperator::createOr(NotCond, TrueVal);
6664 // Selecting between two integer constants?
6665 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6666 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6667 // select C, 1, 0 -> cast C to int
6668 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
6669 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6670 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
6671 // select C, 0, 1 -> cast !C to int
6673 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6674 "not."+CondVal->getName()), SI);
6675 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6678 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
6680 // (x <s 0) ? -1 : 0 -> ashr x, 31
6681 // (x >u 2147483647) ? -1 : 0 -> ashr x, 31
6682 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
6683 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6684 bool CanXForm = false;
6685 if (IC->isSignedPredicate())
6686 CanXForm = CmpCst->isNullValue() &&
6687 IC->getPredicate() == ICmpInst::ICMP_SLT;
6689 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6690 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
6691 IC->getPredicate() == ICmpInst::ICMP_UGT;
6695 // The comparison constant and the result are not neccessarily the
6696 // same width. Make an all-ones value by inserting a AShr.
6697 Value *X = IC->getOperand(0);
6698 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6699 Constant *ShAmt = ConstantInt::get(Type::UByteTy, Bits-1);
6700 Instruction *SRA = new ShiftInst(Instruction::AShr, X,
6702 InsertNewInstBefore(SRA, SI);
6704 // Finally, convert to the type of the select RHS. We figure out
6705 // if this requires a SExt, Trunc or BitCast based on the sizes.
6706 Instruction::CastOps opc = Instruction::BitCast;
6707 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6708 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6709 if (SRASize < SISize)
6710 opc = Instruction::SExt;
6711 else if (SRASize > SISize)
6712 opc = Instruction::Trunc;
6713 return CastInst::create(opc, SRA, SI.getType());
6718 // If one of the constants is zero (we know they can't both be) and we
6719 // have a fcmp instruction with zero, and we have an 'and' with the
6720 // non-constant value, eliminate this whole mess. This corresponds to
6721 // cases like this: ((X & 27) ? 27 : 0)
6722 if (TrueValC->isNullValue() || FalseValC->isNullValue())
6723 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6724 cast<Constant>(IC->getOperand(1))->isNullValue())
6725 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6726 if (ICA->getOpcode() == Instruction::And &&
6727 isa<ConstantInt>(ICA->getOperand(1)) &&
6728 (ICA->getOperand(1) == TrueValC ||
6729 ICA->getOperand(1) == FalseValC) &&
6730 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6731 // Okay, now we know that everything is set up, we just don't
6732 // know whether we have a icmp_ne or icmp_eq and whether the
6733 // true or false val is the zero.
6734 bool ShouldNotVal = !TrueValC->isNullValue();
6735 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
6738 V = InsertNewInstBefore(BinaryOperator::create(
6739 Instruction::Xor, V, ICA->getOperand(1)), SI);
6740 return ReplaceInstUsesWith(SI, V);
6745 // See if we are selecting two values based on a comparison of the two values.
6746 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
6747 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
6748 // Transform (X == Y) ? X : Y -> Y
6749 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6750 return ReplaceInstUsesWith(SI, FalseVal);
6751 // Transform (X != Y) ? X : Y -> X
6752 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6753 return ReplaceInstUsesWith(SI, TrueVal);
6754 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6756 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
6757 // Transform (X == Y) ? Y : X -> X
6758 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
6759 return ReplaceInstUsesWith(SI, FalseVal);
6760 // Transform (X != Y) ? Y : X -> Y
6761 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
6762 return ReplaceInstUsesWith(SI, TrueVal);
6763 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6767 // See if we are selecting two values based on a comparison of the two values.
6768 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
6769 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
6770 // Transform (X == Y) ? X : Y -> Y
6771 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6772 return ReplaceInstUsesWith(SI, FalseVal);
6773 // Transform (X != Y) ? X : Y -> X
6774 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6775 return ReplaceInstUsesWith(SI, TrueVal);
6776 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6778 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
6779 // Transform (X == Y) ? Y : X -> X
6780 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
6781 return ReplaceInstUsesWith(SI, FalseVal);
6782 // Transform (X != Y) ? Y : X -> Y
6783 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
6784 return ReplaceInstUsesWith(SI, TrueVal);
6785 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6789 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6790 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6791 if (TI->hasOneUse() && FI->hasOneUse()) {
6792 Instruction *AddOp = 0, *SubOp = 0;
6794 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6795 if (TI->getOpcode() == FI->getOpcode())
6796 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6799 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6800 // even legal for FP.
6801 if (TI->getOpcode() == Instruction::Sub &&
6802 FI->getOpcode() == Instruction::Add) {
6803 AddOp = FI; SubOp = TI;
6804 } else if (FI->getOpcode() == Instruction::Sub &&
6805 TI->getOpcode() == Instruction::Add) {
6806 AddOp = TI; SubOp = FI;
6810 Value *OtherAddOp = 0;
6811 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6812 OtherAddOp = AddOp->getOperand(1);
6813 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6814 OtherAddOp = AddOp->getOperand(0);
6818 // So at this point we know we have (Y -> OtherAddOp):
6819 // select C, (add X, Y), (sub X, Z)
6820 Value *NegVal; // Compute -Z
6821 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
6822 NegVal = ConstantExpr::getNeg(C);
6824 NegVal = InsertNewInstBefore(
6825 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
6828 Value *NewTrueOp = OtherAddOp;
6829 Value *NewFalseOp = NegVal;
6831 std::swap(NewTrueOp, NewFalseOp);
6832 Instruction *NewSel =
6833 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
6835 NewSel = InsertNewInstBefore(NewSel, SI);
6836 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
6841 // See if we can fold the select into one of our operands.
6842 if (SI.getType()->isInteger()) {
6843 // See the comment above GetSelectFoldableOperands for a description of the
6844 // transformation we are doing here.
6845 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
6846 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
6847 !isa<Constant>(FalseVal))
6848 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
6849 unsigned OpToFold = 0;
6850 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
6852 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
6857 Constant *C = GetSelectFoldableConstant(TVI);
6858 std::string Name = TVI->getName(); TVI->setName("");
6859 Instruction *NewSel =
6860 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
6862 InsertNewInstBefore(NewSel, SI);
6863 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
6864 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
6865 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
6866 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
6868 assert(0 && "Unknown instruction!!");
6873 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
6874 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
6875 !isa<Constant>(TrueVal))
6876 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
6877 unsigned OpToFold = 0;
6878 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
6880 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
6885 Constant *C = GetSelectFoldableConstant(FVI);
6886 std::string Name = FVI->getName(); FVI->setName("");
6887 Instruction *NewSel =
6888 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
6890 InsertNewInstBefore(NewSel, SI);
6891 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
6892 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
6893 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
6894 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
6896 assert(0 && "Unknown instruction!!");
6902 if (BinaryOperator::isNot(CondVal)) {
6903 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
6904 SI.setOperand(1, FalseVal);
6905 SI.setOperand(2, TrueVal);
6912 /// GetKnownAlignment - If the specified pointer has an alignment that we can
6913 /// determine, return it, otherwise return 0.
6914 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
6915 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
6916 unsigned Align = GV->getAlignment();
6917 if (Align == 0 && TD)
6918 Align = TD->getTypeAlignment(GV->getType()->getElementType());
6920 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
6921 unsigned Align = AI->getAlignment();
6922 if (Align == 0 && TD) {
6923 if (isa<AllocaInst>(AI))
6924 Align = TD->getTypeAlignment(AI->getType()->getElementType());
6925 else if (isa<MallocInst>(AI)) {
6926 // Malloc returns maximally aligned memory.
6927 Align = TD->getTypeAlignment(AI->getType()->getElementType());
6928 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::DoubleTy));
6929 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::LongTy));
6933 } else if (isa<BitCastInst>(V) ||
6934 (isa<ConstantExpr>(V) &&
6935 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
6936 User *CI = cast<User>(V);
6937 if (isa<PointerType>(CI->getOperand(0)->getType()))
6938 return GetKnownAlignment(CI->getOperand(0), TD);
6940 } else if (isa<GetElementPtrInst>(V) ||
6941 (isa<ConstantExpr>(V) &&
6942 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
6943 User *GEPI = cast<User>(V);
6944 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
6945 if (BaseAlignment == 0) return 0;
6947 // If all indexes are zero, it is just the alignment of the base pointer.
6948 bool AllZeroOperands = true;
6949 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
6950 if (!isa<Constant>(GEPI->getOperand(i)) ||
6951 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
6952 AllZeroOperands = false;
6955 if (AllZeroOperands)
6956 return BaseAlignment;
6958 // Otherwise, if the base alignment is >= the alignment we expect for the
6959 // base pointer type, then we know that the resultant pointer is aligned at
6960 // least as much as its type requires.
6963 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
6964 if (TD->getTypeAlignment(cast<PointerType>(BasePtrTy)->getElementType())
6966 const Type *GEPTy = GEPI->getType();
6967 return TD->getTypeAlignment(cast<PointerType>(GEPTy)->getElementType());
6975 /// visitCallInst - CallInst simplification. This mostly only handles folding
6976 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
6977 /// the heavy lifting.
6979 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
6980 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
6981 if (!II) return visitCallSite(&CI);
6983 // Intrinsics cannot occur in an invoke, so handle them here instead of in
6985 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
6986 bool Changed = false;
6988 // memmove/cpy/set of zero bytes is a noop.
6989 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
6990 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
6992 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
6993 if (CI->getZExtValue() == 1) {
6994 // Replace the instruction with just byte operations. We would
6995 // transform other cases to loads/stores, but we don't know if
6996 // alignment is sufficient.
7000 // If we have a memmove and the source operation is a constant global,
7001 // then the source and dest pointers can't alias, so we can change this
7002 // into a call to memcpy.
7003 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7004 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7005 if (GVSrc->isConstant()) {
7006 Module *M = CI.getParent()->getParent()->getParent();
7008 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7010 Name = "llvm.memcpy.i32";
7012 Name = "llvm.memcpy.i64";
7013 Function *MemCpy = M->getOrInsertFunction(Name,
7014 CI.getCalledFunction()->getFunctionType());
7015 CI.setOperand(0, MemCpy);
7020 // If we can determine a pointer alignment that is bigger than currently
7021 // set, update the alignment.
7022 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7023 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7024 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7025 unsigned Align = std::min(Alignment1, Alignment2);
7026 if (MI->getAlignment()->getZExtValue() < Align) {
7027 MI->setAlignment(ConstantInt::get(Type::UIntTy, Align));
7030 } else if (isa<MemSetInst>(MI)) {
7031 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7032 if (MI->getAlignment()->getZExtValue() < Alignment) {
7033 MI->setAlignment(ConstantInt::get(Type::UIntTy, Alignment));
7038 if (Changed) return II;
7040 switch (II->getIntrinsicID()) {
7042 case Intrinsic::ppc_altivec_lvx:
7043 case Intrinsic::ppc_altivec_lvxl:
7044 case Intrinsic::x86_sse_loadu_ps:
7045 case Intrinsic::x86_sse2_loadu_pd:
7046 case Intrinsic::x86_sse2_loadu_dq:
7047 // Turn PPC lvx -> load if the pointer is known aligned.
7048 // Turn X86 loadups -> load if the pointer is known aligned.
7049 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7050 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7051 PointerType::get(II->getType()), CI);
7052 return new LoadInst(Ptr);
7055 case Intrinsic::ppc_altivec_stvx:
7056 case Intrinsic::ppc_altivec_stvxl:
7057 // Turn stvx -> store if the pointer is known aligned.
7058 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7059 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7060 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7062 return new StoreInst(II->getOperand(1), Ptr);
7065 case Intrinsic::x86_sse_storeu_ps:
7066 case Intrinsic::x86_sse2_storeu_pd:
7067 case Intrinsic::x86_sse2_storeu_dq:
7068 case Intrinsic::x86_sse2_storel_dq:
7069 // Turn X86 storeu -> store if the pointer is known aligned.
7070 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7071 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7072 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7074 return new StoreInst(II->getOperand(2), Ptr);
7078 case Intrinsic::x86_sse_cvttss2si: {
7079 // These intrinsics only demands the 0th element of its input vector. If
7080 // we can simplify the input based on that, do so now.
7082 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7084 II->setOperand(1, V);
7090 case Intrinsic::ppc_altivec_vperm:
7091 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7092 if (ConstantPacked *Mask = dyn_cast<ConstantPacked>(II->getOperand(3))) {
7093 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7095 // Check that all of the elements are integer constants or undefs.
7096 bool AllEltsOk = true;
7097 for (unsigned i = 0; i != 16; ++i) {
7098 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7099 !isa<UndefValue>(Mask->getOperand(i))) {
7106 // Cast the input vectors to byte vectors.
7107 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7108 II->getOperand(1), Mask->getType(), CI);
7109 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7110 II->getOperand(2), Mask->getType(), CI);
7111 Value *Result = UndefValue::get(Op0->getType());
7113 // Only extract each element once.
7114 Value *ExtractedElts[32];
7115 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7117 for (unsigned i = 0; i != 16; ++i) {
7118 if (isa<UndefValue>(Mask->getOperand(i)))
7120 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7121 Idx &= 31; // Match the hardware behavior.
7123 if (ExtractedElts[Idx] == 0) {
7125 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7126 InsertNewInstBefore(Elt, CI);
7127 ExtractedElts[Idx] = Elt;
7130 // Insert this value into the result vector.
7131 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7132 InsertNewInstBefore(cast<Instruction>(Result), CI);
7134 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7139 case Intrinsic::stackrestore: {
7140 // If the save is right next to the restore, remove the restore. This can
7141 // happen when variable allocas are DCE'd.
7142 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7143 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7144 BasicBlock::iterator BI = SS;
7146 return EraseInstFromFunction(CI);
7150 // If the stack restore is in a return/unwind block and if there are no
7151 // allocas or calls between the restore and the return, nuke the restore.
7152 TerminatorInst *TI = II->getParent()->getTerminator();
7153 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7154 BasicBlock::iterator BI = II;
7155 bool CannotRemove = false;
7156 for (++BI; &*BI != TI; ++BI) {
7157 if (isa<AllocaInst>(BI) ||
7158 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7159 CannotRemove = true;
7164 return EraseInstFromFunction(CI);
7171 return visitCallSite(II);
7174 // InvokeInst simplification
7176 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7177 return visitCallSite(&II);
7180 // visitCallSite - Improvements for call and invoke instructions.
7182 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7183 bool Changed = false;
7185 // If the callee is a constexpr cast of a function, attempt to move the cast
7186 // to the arguments of the call/invoke.
7187 if (transformConstExprCastCall(CS)) return 0;
7189 Value *Callee = CS.getCalledValue();
7191 if (Function *CalleeF = dyn_cast<Function>(Callee))
7192 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7193 Instruction *OldCall = CS.getInstruction();
7194 // If the call and callee calling conventions don't match, this call must
7195 // be unreachable, as the call is undefined.
7196 new StoreInst(ConstantBool::getTrue(),
7197 UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
7198 if (!OldCall->use_empty())
7199 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7200 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7201 return EraseInstFromFunction(*OldCall);
7205 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7206 // This instruction is not reachable, just remove it. We insert a store to
7207 // undef so that we know that this code is not reachable, despite the fact
7208 // that we can't modify the CFG here.
7209 new StoreInst(ConstantBool::getTrue(),
7210 UndefValue::get(PointerType::get(Type::BoolTy)),
7211 CS.getInstruction());
7213 if (!CS.getInstruction()->use_empty())
7214 CS.getInstruction()->
7215 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7217 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7218 // Don't break the CFG, insert a dummy cond branch.
7219 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7220 ConstantBool::getTrue(), II);
7222 return EraseInstFromFunction(*CS.getInstruction());
7225 const PointerType *PTy = cast<PointerType>(Callee->getType());
7226 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7227 if (FTy->isVarArg()) {
7228 // See if we can optimize any arguments passed through the varargs area of
7230 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7231 E = CS.arg_end(); I != E; ++I)
7232 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7233 // If this cast does not effect the value passed through the varargs
7234 // area, we can eliminate the use of the cast.
7235 Value *Op = CI->getOperand(0);
7236 if (CI->isLosslessCast()) {
7243 return Changed ? CS.getInstruction() : 0;
7246 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7247 // attempt to move the cast to the arguments of the call/invoke.
7249 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7250 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7251 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7252 if (CE->getOpcode() != Instruction::BitCast ||
7253 !isa<Function>(CE->getOperand(0)))
7255 Function *Callee = cast<Function>(CE->getOperand(0));
7256 Instruction *Caller = CS.getInstruction();
7258 // Okay, this is a cast from a function to a different type. Unless doing so
7259 // would cause a type conversion of one of our arguments, change this call to
7260 // be a direct call with arguments casted to the appropriate types.
7262 const FunctionType *FT = Callee->getFunctionType();
7263 const Type *OldRetTy = Caller->getType();
7265 // Check to see if we are changing the return type...
7266 if (OldRetTy != FT->getReturnType()) {
7267 if (Callee->isExternal() &&
7268 !Caller->use_empty() &&
7269 !(OldRetTy->canLosslesslyBitCastTo(FT->getReturnType()) ||
7270 (isa<PointerType>(FT->getReturnType()) &&
7271 TD->getIntPtrType()->canLosslesslyBitCastTo(OldRetTy)))
7273 return false; // Cannot transform this return value...
7275 // If the callsite is an invoke instruction, and the return value is used by
7276 // a PHI node in a successor, we cannot change the return type of the call
7277 // because there is no place to put the cast instruction (without breaking
7278 // the critical edge). Bail out in this case.
7279 if (!Caller->use_empty())
7280 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7281 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7283 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7284 if (PN->getParent() == II->getNormalDest() ||
7285 PN->getParent() == II->getUnwindDest())
7289 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7290 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7292 CallSite::arg_iterator AI = CS.arg_begin();
7293 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7294 const Type *ParamTy = FT->getParamType(i);
7295 const Type *ActTy = (*AI)->getType();
7296 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7297 //Either we can cast directly, or we can upconvert the argument
7298 bool isConvertible = ActTy->canLosslesslyBitCastTo(ParamTy) ||
7299 (ParamTy->isIntegral() && ActTy->isIntegral() &&
7300 ParamTy->isSigned() == ActTy->isSigned() &&
7301 ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize()) ||
7302 (c && ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize() &&
7303 c->getSExtValue() > 0);
7304 if (Callee->isExternal() && !isConvertible) return false;
7307 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7308 Callee->isExternal())
7309 return false; // Do not delete arguments unless we have a function body...
7311 // Okay, we decided that this is a safe thing to do: go ahead and start
7312 // inserting cast instructions as necessary...
7313 std::vector<Value*> Args;
7314 Args.reserve(NumActualArgs);
7316 AI = CS.arg_begin();
7317 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7318 const Type *ParamTy = FT->getParamType(i);
7319 if ((*AI)->getType() == ParamTy) {
7320 Args.push_back(*AI);
7322 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7323 (*AI)->getType()->isSigned(), ParamTy, ParamTy->isSigned());
7324 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7325 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7329 // If the function takes more arguments than the call was taking, add them
7331 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7332 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7334 // If we are removing arguments to the function, emit an obnoxious warning...
7335 if (FT->getNumParams() < NumActualArgs)
7336 if (!FT->isVarArg()) {
7337 cerr << "WARNING: While resolving call to function '"
7338 << Callee->getName() << "' arguments were dropped!\n";
7340 // Add all of the arguments in their promoted form to the arg list...
7341 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7342 const Type *PTy = getPromotedType((*AI)->getType());
7343 if (PTy != (*AI)->getType()) {
7344 // Must promote to pass through va_arg area!
7345 Instruction::CastOps opcode = CastInst::getCastOpcode(
7346 *AI, (*AI)->getType()->isSigned(), PTy, PTy->isSigned());
7347 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7348 InsertNewInstBefore(Cast, *Caller);
7349 Args.push_back(Cast);
7351 Args.push_back(*AI);
7356 if (FT->getReturnType() == Type::VoidTy)
7357 Caller->setName(""); // Void type should not have a name...
7360 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7361 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7362 Args, Caller->getName(), Caller);
7363 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7365 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
7366 if (cast<CallInst>(Caller)->isTailCall())
7367 cast<CallInst>(NC)->setTailCall();
7368 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7371 // Insert a cast of the return type as necessary...
7373 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7374 if (NV->getType() != Type::VoidTy) {
7375 const Type *CallerTy = Caller->getType();
7376 Instruction::CastOps opcode = CastInst::getCastOpcode(
7377 NC, NC->getType()->isSigned(), CallerTy, CallerTy->isSigned());
7378 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7380 // If this is an invoke instruction, we should insert it after the first
7381 // non-phi, instruction in the normal successor block.
7382 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7383 BasicBlock::iterator I = II->getNormalDest()->begin();
7384 while (isa<PHINode>(I)) ++I;
7385 InsertNewInstBefore(NC, *I);
7387 // Otherwise, it's a call, just insert cast right after the call instr
7388 InsertNewInstBefore(NC, *Caller);
7390 AddUsersToWorkList(*Caller);
7392 NV = UndefValue::get(Caller->getType());
7396 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7397 Caller->replaceAllUsesWith(NV);
7398 Caller->getParent()->getInstList().erase(Caller);
7399 removeFromWorkList(Caller);
7403 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7404 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7405 /// and a single binop.
7406 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7407 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7408 assert(isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
7409 isa<GetElementPtrInst>(FirstInst) || isa<CmpInst>(FirstInst));
7410 unsigned Opc = FirstInst->getOpcode();
7411 Value *LHSVal = FirstInst->getOperand(0);
7412 Value *RHSVal = FirstInst->getOperand(1);
7414 const Type *LHSType = LHSVal->getType();
7415 const Type *RHSType = RHSVal->getType();
7417 // Scan to see if all operands are the same opcode, all have one use, and all
7418 // kill their operands (i.e. the operands have one use).
7419 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7420 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7421 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7422 // Verify type of the LHS matches so we don't fold cmp's of different
7423 // types or GEP's with different index types.
7424 I->getOperand(0)->getType() != LHSType ||
7425 I->getOperand(1)->getType() != RHSType)
7428 // If they are CmpInst instructions, check their predicates
7429 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7430 if (cast<CmpInst>(I)->getPredicate() !=
7431 cast<CmpInst>(FirstInst)->getPredicate())
7434 // Keep track of which operand needs a phi node.
7435 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7436 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7439 // Otherwise, this is safe to transform, determine if it is profitable.
7441 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7442 // Indexes are often folded into load/store instructions, so we don't want to
7443 // hide them behind a phi.
7444 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7447 Value *InLHS = FirstInst->getOperand(0);
7448 Value *InRHS = FirstInst->getOperand(1);
7449 PHINode *NewLHS = 0, *NewRHS = 0;
7451 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7452 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7453 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7454 InsertNewInstBefore(NewLHS, PN);
7459 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7460 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7461 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7462 InsertNewInstBefore(NewRHS, PN);
7466 // Add all operands to the new PHIs.
7467 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7469 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7470 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7473 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7474 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7478 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7479 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7480 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7481 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
7483 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FirstInst))
7484 return new ShiftInst(SI->getOpcode(), LHSVal, RHSVal);
7486 assert(isa<GetElementPtrInst>(FirstInst));
7487 return new GetElementPtrInst(LHSVal, RHSVal);
7491 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7492 /// of the block that defines it. This means that it must be obvious the value
7493 /// of the load is not changed from the point of the load to the end of the
7495 static bool isSafeToSinkLoad(LoadInst *L) {
7496 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7498 for (++BBI; BBI != E; ++BBI)
7499 if (BBI->mayWriteToMemory())
7505 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7506 // operator and they all are only used by the PHI, PHI together their
7507 // inputs, and do the operation once, to the result of the PHI.
7508 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7509 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7511 // Scan the instruction, looking for input operations that can be folded away.
7512 // If all input operands to the phi are the same instruction (e.g. a cast from
7513 // the same type or "+42") we can pull the operation through the PHI, reducing
7514 // code size and simplifying code.
7515 Constant *ConstantOp = 0;
7516 const Type *CastSrcTy = 0;
7517 bool isVolatile = false;
7518 if (isa<CastInst>(FirstInst)) {
7519 CastSrcTy = FirstInst->getOperand(0)->getType();
7520 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
7521 isa<CmpInst>(FirstInst)) {
7522 // Can fold binop, compare or shift here if the RHS is a constant,
7523 // otherwise call FoldPHIArgBinOpIntoPHI.
7524 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7525 if (ConstantOp == 0)
7526 return FoldPHIArgBinOpIntoPHI(PN);
7527 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7528 isVolatile = LI->isVolatile();
7529 // We can't sink the load if the loaded value could be modified between the
7530 // load and the PHI.
7531 if (LI->getParent() != PN.getIncomingBlock(0) ||
7532 !isSafeToSinkLoad(LI))
7534 } else if (isa<GetElementPtrInst>(FirstInst)) {
7535 if (FirstInst->getNumOperands() == 2)
7536 return FoldPHIArgBinOpIntoPHI(PN);
7537 // Can't handle general GEPs yet.
7540 return 0; // Cannot fold this operation.
7543 // Check to see if all arguments are the same operation.
7544 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7545 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7546 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7547 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
7550 if (I->getOperand(0)->getType() != CastSrcTy)
7551 return 0; // Cast operation must match.
7552 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7553 // We can't sink the load if the loaded value could be modified between
7554 // the load and the PHI.
7555 if (LI->isVolatile() != isVolatile ||
7556 LI->getParent() != PN.getIncomingBlock(i) ||
7557 !isSafeToSinkLoad(LI))
7559 } else if (I->getOperand(1) != ConstantOp) {
7564 // Okay, they are all the same operation. Create a new PHI node of the
7565 // correct type, and PHI together all of the LHS's of the instructions.
7566 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7567 PN.getName()+".in");
7568 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7570 Value *InVal = FirstInst->getOperand(0);
7571 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7573 // Add all operands to the new PHI.
7574 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7575 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7576 if (NewInVal != InVal)
7578 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7583 // The new PHI unions all of the same values together. This is really
7584 // common, so we handle it intelligently here for compile-time speed.
7588 InsertNewInstBefore(NewPN, PN);
7592 // Insert and return the new operation.
7593 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7594 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7595 else if (isa<LoadInst>(FirstInst))
7596 return new LoadInst(PhiVal, "", isVolatile);
7597 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7598 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7599 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
7600 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
7601 PhiVal, ConstantOp);
7603 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
7604 PhiVal, ConstantOp);
7607 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7609 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7610 if (PN->use_empty()) return true;
7611 if (!PN->hasOneUse()) return false;
7613 // Remember this node, and if we find the cycle, return.
7614 if (!PotentiallyDeadPHIs.insert(PN).second)
7617 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7618 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7623 // PHINode simplification
7625 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7626 // If LCSSA is around, don't mess with Phi nodes
7627 if (mustPreserveAnalysisID(LCSSAID)) return 0;
7629 if (Value *V = PN.hasConstantValue())
7630 return ReplaceInstUsesWith(PN, V);
7632 // If all PHI operands are the same operation, pull them through the PHI,
7633 // reducing code size.
7634 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7635 PN.getIncomingValue(0)->hasOneUse())
7636 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7639 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7640 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7641 // PHI)... break the cycle.
7643 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
7644 std::set<PHINode*> PotentiallyDeadPHIs;
7645 PotentiallyDeadPHIs.insert(&PN);
7646 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7647 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7653 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7654 Instruction *InsertPoint,
7656 unsigned PtrSize = DTy->getPrimitiveSize();
7657 unsigned VTySize = V->getType()->getPrimitiveSize();
7658 // We must cast correctly to the pointer type. Ensure that we
7659 // sign extend the integer value if it is smaller as this is
7660 // used for address computation.
7661 Instruction::CastOps opcode =
7662 (VTySize < PtrSize ? Instruction::SExt :
7663 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7664 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7668 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7669 Value *PtrOp = GEP.getOperand(0);
7670 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7671 // If so, eliminate the noop.
7672 if (GEP.getNumOperands() == 1)
7673 return ReplaceInstUsesWith(GEP, PtrOp);
7675 if (isa<UndefValue>(GEP.getOperand(0)))
7676 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7678 bool HasZeroPointerIndex = false;
7679 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7680 HasZeroPointerIndex = C->isNullValue();
7682 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7683 return ReplaceInstUsesWith(GEP, PtrOp);
7685 // Eliminate unneeded casts for indices.
7686 bool MadeChange = false;
7687 gep_type_iterator GTI = gep_type_begin(GEP);
7688 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7689 if (isa<SequentialType>(*GTI)) {
7690 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7691 Value *Src = CI->getOperand(0);
7692 const Type *SrcTy = Src->getType();
7693 const Type *DestTy = CI->getType();
7694 if (Src->getType()->isInteger()) {
7695 if (SrcTy->getPrimitiveSizeInBits() ==
7696 DestTy->getPrimitiveSizeInBits()) {
7697 // We can always eliminate a cast from ulong or long to the other.
7698 // We can always eliminate a cast from uint to int or the other on
7699 // 32-bit pointer platforms.
7700 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
7702 GEP.setOperand(i, Src);
7704 } else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
7705 SrcTy->getPrimitiveSize() == 4) {
7706 // We can eliminate a cast from [u]int to [u]long iff the target
7707 // is a 32-bit pointer target.
7708 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7710 GEP.setOperand(i, Src);
7715 // If we are using a wider index than needed for this platform, shrink it
7716 // to what we need. If the incoming value needs a cast instruction,
7717 // insert it. This explicit cast can make subsequent optimizations more
7719 Value *Op = GEP.getOperand(i);
7720 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
7721 if (Constant *C = dyn_cast<Constant>(Op)) {
7722 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
7725 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7727 GEP.setOperand(i, Op);
7730 // If this is a constant idx, make sure to canonicalize it to be a signed
7731 // operand, otherwise CSE and other optimizations are pessimized.
7732 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op))
7733 if (CUI->getType()->isUnsigned()) {
7735 ConstantExpr::getBitCast(CUI, CUI->getType()->getSignedVersion()));
7739 if (MadeChange) return &GEP;
7741 // Combine Indices - If the source pointer to this getelementptr instruction
7742 // is a getelementptr instruction, combine the indices of the two
7743 // getelementptr instructions into a single instruction.
7745 std::vector<Value*> SrcGEPOperands;
7746 if (User *Src = dyn_castGetElementPtr(PtrOp))
7747 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
7749 if (!SrcGEPOperands.empty()) {
7750 // Note that if our source is a gep chain itself that we wait for that
7751 // chain to be resolved before we perform this transformation. This
7752 // avoids us creating a TON of code in some cases.
7754 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7755 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7756 return 0; // Wait until our source is folded to completion.
7758 std::vector<Value *> Indices;
7760 // Find out whether the last index in the source GEP is a sequential idx.
7761 bool EndsWithSequential = false;
7762 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7763 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7764 EndsWithSequential = !isa<StructType>(*I);
7766 // Can we combine the two pointer arithmetics offsets?
7767 if (EndsWithSequential) {
7768 // Replace: gep (gep %P, long B), long A, ...
7769 // With: T = long A+B; gep %P, T, ...
7771 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7772 if (SO1 == Constant::getNullValue(SO1->getType())) {
7774 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7777 // If they aren't the same type, convert both to an integer of the
7778 // target's pointer size.
7779 if (SO1->getType() != GO1->getType()) {
7780 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7781 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
7782 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7783 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
7785 unsigned PS = TD->getPointerSize();
7786 if (SO1->getType()->getPrimitiveSize() == PS) {
7787 // Convert GO1 to SO1's type.
7788 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
7790 } else if (GO1->getType()->getPrimitiveSize() == PS) {
7791 // Convert SO1 to GO1's type.
7792 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
7794 const Type *PT = TD->getIntPtrType();
7795 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
7796 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
7800 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7801 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7803 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7804 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7808 // Recycle the GEP we already have if possible.
7809 if (SrcGEPOperands.size() == 2) {
7810 GEP.setOperand(0, SrcGEPOperands[0]);
7811 GEP.setOperand(1, Sum);
7814 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7815 SrcGEPOperands.end()-1);
7816 Indices.push_back(Sum);
7817 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7819 } else if (isa<Constant>(*GEP.idx_begin()) &&
7820 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7821 SrcGEPOperands.size() != 1) {
7822 // Otherwise we can do the fold if the first index of the GEP is a zero
7823 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7824 SrcGEPOperands.end());
7825 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7828 if (!Indices.empty())
7829 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
7831 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7832 // GEP of global variable. If all of the indices for this GEP are
7833 // constants, we can promote this to a constexpr instead of an instruction.
7835 // Scan for nonconstants...
7836 std::vector<Constant*> Indices;
7837 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7838 for (; I != E && isa<Constant>(*I); ++I)
7839 Indices.push_back(cast<Constant>(*I));
7841 if (I == E) { // If they are all constants...
7842 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
7844 // Replace all uses of the GEP with the new constexpr...
7845 return ReplaceInstUsesWith(GEP, CE);
7847 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
7848 if (!isa<PointerType>(X->getType())) {
7849 // Not interesting. Source pointer must be a cast from pointer.
7850 } else if (HasZeroPointerIndex) {
7851 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7852 // into : GEP [10 x ubyte]* X, long 0, ...
7854 // This occurs when the program declares an array extern like "int X[];"
7856 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7857 const PointerType *XTy = cast<PointerType>(X->getType());
7858 if (const ArrayType *XATy =
7859 dyn_cast<ArrayType>(XTy->getElementType()))
7860 if (const ArrayType *CATy =
7861 dyn_cast<ArrayType>(CPTy->getElementType()))
7862 if (CATy->getElementType() == XATy->getElementType()) {
7863 // At this point, we know that the cast source type is a pointer
7864 // to an array of the same type as the destination pointer
7865 // array. Because the array type is never stepped over (there
7866 // is a leading zero) we can fold the cast into this GEP.
7867 GEP.setOperand(0, X);
7870 } else if (GEP.getNumOperands() == 2) {
7871 // Transform things like:
7872 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7873 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7874 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7875 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7876 if (isa<ArrayType>(SrcElTy) &&
7877 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7878 TD->getTypeSize(ResElTy)) {
7879 Value *V = InsertNewInstBefore(
7880 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
7881 GEP.getOperand(1), GEP.getName()), GEP);
7882 // V and GEP are both pointer types --> BitCast
7883 return new BitCastInst(V, GEP.getType());
7886 // Transform things like:
7887 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7888 // (where tmp = 8*tmp2) into:
7889 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7891 if (isa<ArrayType>(SrcElTy) &&
7892 (ResElTy == Type::SByteTy || ResElTy == Type::UByteTy)) {
7893 uint64_t ArrayEltSize =
7894 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7896 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7897 // allow either a mul, shift, or constant here.
7899 ConstantInt *Scale = 0;
7900 if (ArrayEltSize == 1) {
7901 NewIdx = GEP.getOperand(1);
7902 Scale = ConstantInt::get(NewIdx->getType(), 1);
7903 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7904 NewIdx = ConstantInt::get(CI->getType(), 1);
7906 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7907 if (Inst->getOpcode() == Instruction::Shl &&
7908 isa<ConstantInt>(Inst->getOperand(1))) {
7910 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
7911 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7912 NewIdx = Inst->getOperand(0);
7913 } else if (Inst->getOpcode() == Instruction::Mul &&
7914 isa<ConstantInt>(Inst->getOperand(1))) {
7915 Scale = cast<ConstantInt>(Inst->getOperand(1));
7916 NewIdx = Inst->getOperand(0);
7920 // If the index will be to exactly the right offset with the scale taken
7921 // out, perform the transformation.
7922 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
7923 if (isa<ConstantInt>(Scale))
7924 Scale = ConstantInt::get(Scale->getType(),
7925 Scale->getZExtValue() / ArrayEltSize);
7926 if (Scale->getZExtValue() != 1) {
7927 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
7929 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
7930 NewIdx = InsertNewInstBefore(Sc, GEP);
7933 // Insert the new GEP instruction.
7934 Instruction *NewGEP =
7935 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
7936 NewIdx, GEP.getName());
7937 NewGEP = InsertNewInstBefore(NewGEP, GEP);
7938 // The NewGEP must be pointer typed, so must the old one -> BitCast
7939 return new BitCastInst(NewGEP, GEP.getType());
7948 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
7949 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
7950 if (AI.isArrayAllocation()) // Check C != 1
7951 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7953 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7954 AllocationInst *New = 0;
7956 // Create and insert the replacement instruction...
7957 if (isa<MallocInst>(AI))
7958 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
7960 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7961 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
7964 InsertNewInstBefore(New, AI);
7966 // Scan to the end of the allocation instructions, to skip over a block of
7967 // allocas if possible...
7969 BasicBlock::iterator It = New;
7970 while (isa<AllocationInst>(*It)) ++It;
7972 // Now that I is pointing to the first non-allocation-inst in the block,
7973 // insert our getelementptr instruction...
7975 Value *NullIdx = Constant::getNullValue(Type::IntTy);
7976 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
7977 New->getName()+".sub", It);
7979 // Now make everything use the getelementptr instead of the original
7981 return ReplaceInstUsesWith(AI, V);
7982 } else if (isa<UndefValue>(AI.getArraySize())) {
7983 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7986 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
7987 // Note that we only do this for alloca's, because malloc should allocate and
7988 // return a unique pointer, even for a zero byte allocation.
7989 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
7990 TD->getTypeSize(AI.getAllocatedType()) == 0)
7991 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7996 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
7997 Value *Op = FI.getOperand(0);
7999 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8000 if (CastInst *CI = dyn_cast<CastInst>(Op))
8001 if (isa<PointerType>(CI->getOperand(0)->getType())) {
8002 FI.setOperand(0, CI->getOperand(0));
8006 // free undef -> unreachable.
8007 if (isa<UndefValue>(Op)) {
8008 // Insert a new store to null because we cannot modify the CFG here.
8009 new StoreInst(ConstantBool::getTrue(),
8010 UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
8011 return EraseInstFromFunction(FI);
8014 // If we have 'free null' delete the instruction. This can happen in stl code
8015 // when lots of inlining happens.
8016 if (isa<ConstantPointerNull>(Op))
8017 return EraseInstFromFunction(FI);
8023 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8024 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8025 User *CI = cast<User>(LI.getOperand(0));
8026 Value *CastOp = CI->getOperand(0);
8028 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8029 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8030 const Type *SrcPTy = SrcTy->getElementType();
8032 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8033 isa<PackedType>(DestPTy)) {
8034 // If the source is an array, the code below will not succeed. Check to
8035 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8037 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8038 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8039 if (ASrcTy->getNumElements() != 0) {
8040 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
8041 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
8042 SrcTy = cast<PointerType>(CastOp->getType());
8043 SrcPTy = SrcTy->getElementType();
8046 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8047 isa<PackedType>(SrcPTy)) &&
8048 // Do not allow turning this into a load of an integer, which is then
8049 // casted to a pointer, this pessimizes pointer analysis a lot.
8050 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8051 IC.getTargetData().getTypeSize(SrcPTy) ==
8052 IC.getTargetData().getTypeSize(DestPTy)) {
8054 // Okay, we are casting from one integer or pointer type to another of
8055 // the same size. Instead of casting the pointer before the load, cast
8056 // the result of the loaded value.
8057 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8059 LI.isVolatile()),LI);
8060 // Now cast the result of the load.
8061 return new BitCastInst(NewLoad, LI.getType());
8068 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8069 /// from this value cannot trap. If it is not obviously safe to load from the
8070 /// specified pointer, we do a quick local scan of the basic block containing
8071 /// ScanFrom, to determine if the address is already accessed.
8072 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8073 // If it is an alloca or global variable, it is always safe to load from.
8074 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8076 // Otherwise, be a little bit agressive by scanning the local block where we
8077 // want to check to see if the pointer is already being loaded or stored
8078 // from/to. If so, the previous load or store would have already trapped,
8079 // so there is no harm doing an extra load (also, CSE will later eliminate
8080 // the load entirely).
8081 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8086 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8087 if (LI->getOperand(0) == V) return true;
8088 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8089 if (SI->getOperand(1) == V) return true;
8095 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8096 Value *Op = LI.getOperand(0);
8098 // load (cast X) --> cast (load X) iff safe
8099 if (isa<CastInst>(Op))
8100 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8103 // None of the following transforms are legal for volatile loads.
8104 if (LI.isVolatile()) return 0;
8106 if (&LI.getParent()->front() != &LI) {
8107 BasicBlock::iterator BBI = &LI; --BBI;
8108 // If the instruction immediately before this is a store to the same
8109 // address, do a simple form of store->load forwarding.
8110 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8111 if (SI->getOperand(1) == LI.getOperand(0))
8112 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8113 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8114 if (LIB->getOperand(0) == LI.getOperand(0))
8115 return ReplaceInstUsesWith(LI, LIB);
8118 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8119 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
8120 isa<UndefValue>(GEPI->getOperand(0))) {
8121 // Insert a new store to null instruction before the load to indicate
8122 // that this code is not reachable. We do this instead of inserting
8123 // an unreachable instruction directly because we cannot modify the
8125 new StoreInst(UndefValue::get(LI.getType()),
8126 Constant::getNullValue(Op->getType()), &LI);
8127 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8130 if (Constant *C = dyn_cast<Constant>(Op)) {
8131 // load null/undef -> undef
8132 if ((C->isNullValue() || isa<UndefValue>(C))) {
8133 // Insert a new store to null instruction before the load to indicate that
8134 // this code is not reachable. We do this instead of inserting an
8135 // unreachable instruction directly because we cannot modify the CFG.
8136 new StoreInst(UndefValue::get(LI.getType()),
8137 Constant::getNullValue(Op->getType()), &LI);
8138 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8141 // Instcombine load (constant global) into the value loaded.
8142 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8143 if (GV->isConstant() && !GV->isExternal())
8144 return ReplaceInstUsesWith(LI, GV->getInitializer());
8146 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8147 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8148 if (CE->getOpcode() == Instruction::GetElementPtr) {
8149 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8150 if (GV->isConstant() && !GV->isExternal())
8152 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8153 return ReplaceInstUsesWith(LI, V);
8154 if (CE->getOperand(0)->isNullValue()) {
8155 // Insert a new store to null instruction before the load to indicate
8156 // that this code is not reachable. We do this instead of inserting
8157 // an unreachable instruction directly because we cannot modify the
8159 new StoreInst(UndefValue::get(LI.getType()),
8160 Constant::getNullValue(Op->getType()), &LI);
8161 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8164 } else if (CE->isCast()) {
8165 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8170 if (Op->hasOneUse()) {
8171 // Change select and PHI nodes to select values instead of addresses: this
8172 // helps alias analysis out a lot, allows many others simplifications, and
8173 // exposes redundancy in the code.
8175 // Note that we cannot do the transformation unless we know that the
8176 // introduced loads cannot trap! Something like this is valid as long as
8177 // the condition is always false: load (select bool %C, int* null, int* %G),
8178 // but it would not be valid if we transformed it to load from null
8181 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8182 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8183 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8184 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8185 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8186 SI->getOperand(1)->getName()+".val"), LI);
8187 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8188 SI->getOperand(2)->getName()+".val"), LI);
8189 return new SelectInst(SI->getCondition(), V1, V2);
8192 // load (select (cond, null, P)) -> load P
8193 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8194 if (C->isNullValue()) {
8195 LI.setOperand(0, SI->getOperand(2));
8199 // load (select (cond, P, null)) -> load P
8200 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8201 if (C->isNullValue()) {
8202 LI.setOperand(0, SI->getOperand(1));
8210 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
8212 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8213 User *CI = cast<User>(SI.getOperand(1));
8214 Value *CastOp = CI->getOperand(0);
8216 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8217 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8218 const Type *SrcPTy = SrcTy->getElementType();
8220 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8221 // If the source is an array, the code below will not succeed. Check to
8222 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8224 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8225 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8226 if (ASrcTy->getNumElements() != 0) {
8227 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
8228 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
8229 SrcTy = cast<PointerType>(CastOp->getType());
8230 SrcPTy = SrcTy->getElementType();
8233 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8234 IC.getTargetData().getTypeSize(SrcPTy) ==
8235 IC.getTargetData().getTypeSize(DestPTy)) {
8237 // Okay, we are casting from one integer or pointer type to another of
8238 // the same size. Instead of casting the pointer before the store, cast
8239 // the value to be stored.
8241 Instruction::CastOps opcode = Instruction::BitCast;
8242 Value *SIOp0 = SI.getOperand(0);
8243 if (isa<PointerType>(SrcPTy)) {
8244 if (SIOp0->getType()->isIntegral())
8245 opcode = Instruction::IntToPtr;
8246 } else if (SrcPTy->isIntegral()) {
8247 if (isa<PointerType>(SIOp0->getType()))
8248 opcode = Instruction::PtrToInt;
8250 if (Constant *C = dyn_cast<Constant>(SIOp0))
8251 NewCast = ConstantExpr::getCast(opcode, C, SrcPTy);
8253 NewCast = IC.InsertNewInstBefore(
8254 CastInst::create(opcode, SIOp0, SrcPTy, SIOp0->getName()+".c"), SI);
8255 return new StoreInst(NewCast, CastOp);
8262 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8263 Value *Val = SI.getOperand(0);
8264 Value *Ptr = SI.getOperand(1);
8266 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8267 EraseInstFromFunction(SI);
8272 // Do really simple DSE, to catch cases where there are several consequtive
8273 // stores to the same location, separated by a few arithmetic operations. This
8274 // situation often occurs with bitfield accesses.
8275 BasicBlock::iterator BBI = &SI;
8276 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8280 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8281 // Prev store isn't volatile, and stores to the same location?
8282 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8285 EraseInstFromFunction(*PrevSI);
8291 // If this is a load, we have to stop. However, if the loaded value is from
8292 // the pointer we're loading and is producing the pointer we're storing,
8293 // then *this* store is dead (X = load P; store X -> P).
8294 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8295 if (LI == Val && LI->getOperand(0) == Ptr) {
8296 EraseInstFromFunction(SI);
8300 // Otherwise, this is a load from some other location. Stores before it
8305 // Don't skip over loads or things that can modify memory.
8306 if (BBI->mayWriteToMemory())
8311 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8313 // store X, null -> turns into 'unreachable' in SimplifyCFG
8314 if (isa<ConstantPointerNull>(Ptr)) {
8315 if (!isa<UndefValue>(Val)) {
8316 SI.setOperand(0, UndefValue::get(Val->getType()));
8317 if (Instruction *U = dyn_cast<Instruction>(Val))
8318 WorkList.push_back(U); // Dropped a use.
8321 return 0; // Do not modify these!
8324 // store undef, Ptr -> noop
8325 if (isa<UndefValue>(Val)) {
8326 EraseInstFromFunction(SI);
8331 // If the pointer destination is a cast, see if we can fold the cast into the
8333 if (isa<CastInst>(Ptr))
8334 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8336 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8338 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8342 // If this store is the last instruction in the basic block, and if the block
8343 // ends with an unconditional branch, try to move it to the successor block.
8345 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8346 if (BI->isUnconditional()) {
8347 // Check to see if the successor block has exactly two incoming edges. If
8348 // so, see if the other predecessor contains a store to the same location.
8349 // if so, insert a PHI node (if needed) and move the stores down.
8350 BasicBlock *Dest = BI->getSuccessor(0);
8352 pred_iterator PI = pred_begin(Dest);
8353 BasicBlock *Other = 0;
8354 if (*PI != BI->getParent())
8357 if (PI != pred_end(Dest)) {
8358 if (*PI != BI->getParent())
8363 if (++PI != pred_end(Dest))
8366 if (Other) { // If only one other pred...
8367 BBI = Other->getTerminator();
8368 // Make sure this other block ends in an unconditional branch and that
8369 // there is an instruction before the branch.
8370 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8371 BBI != Other->begin()) {
8373 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8375 // If this instruction is a store to the same location.
8376 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8377 // Okay, we know we can perform this transformation. Insert a PHI
8378 // node now if we need it.
8379 Value *MergedVal = OtherStore->getOperand(0);
8380 if (MergedVal != SI.getOperand(0)) {
8381 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8382 PN->reserveOperandSpace(2);
8383 PN->addIncoming(SI.getOperand(0), SI.getParent());
8384 PN->addIncoming(OtherStore->getOperand(0), Other);
8385 MergedVal = InsertNewInstBefore(PN, Dest->front());
8388 // Advance to a place where it is safe to insert the new store and
8390 BBI = Dest->begin();
8391 while (isa<PHINode>(BBI)) ++BBI;
8392 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8393 OtherStore->isVolatile()), *BBI);
8395 // Nuke the old stores.
8396 EraseInstFromFunction(SI);
8397 EraseInstFromFunction(*OtherStore);
8409 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8410 // Change br (not X), label True, label False to: br X, label False, True
8412 BasicBlock *TrueDest;
8413 BasicBlock *FalseDest;
8414 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8415 !isa<Constant>(X)) {
8416 // Swap Destinations and condition...
8418 BI.setSuccessor(0, FalseDest);
8419 BI.setSuccessor(1, TrueDest);
8423 // Cannonicalize fcmp_one -> fcmp_oeq
8424 FCmpInst::Predicate FPred; Value *Y;
8425 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
8426 TrueDest, FalseDest)))
8427 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
8428 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
8429 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
8430 std::string Name = I->getName(); I->setName("");
8431 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
8432 Value *NewSCC = new FCmpInst(NewPred, X, Y, Name, I);
8433 // Swap Destinations and condition...
8434 BI.setCondition(NewSCC);
8435 BI.setSuccessor(0, FalseDest);
8436 BI.setSuccessor(1, TrueDest);
8437 removeFromWorkList(I);
8438 I->getParent()->getInstList().erase(I);
8439 WorkList.push_back(cast<Instruction>(NewSCC));
8443 // Cannonicalize icmp_ne -> icmp_eq
8444 ICmpInst::Predicate IPred;
8445 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
8446 TrueDest, FalseDest)))
8447 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
8448 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
8449 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
8450 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
8451 std::string Name = I->getName(); I->setName("");
8452 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
8453 Value *NewSCC = new ICmpInst(NewPred, X, Y, Name, I);
8454 // Swap Destinations and condition...
8455 BI.setCondition(NewSCC);
8456 BI.setSuccessor(0, FalseDest);
8457 BI.setSuccessor(1, TrueDest);
8458 removeFromWorkList(I);
8459 I->getParent()->getInstList().erase(I);
8460 WorkList.push_back(cast<Instruction>(NewSCC));
8467 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8468 Value *Cond = SI.getCondition();
8469 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8470 if (I->getOpcode() == Instruction::Add)
8471 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8472 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8473 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8474 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8476 SI.setOperand(0, I->getOperand(0));
8477 WorkList.push_back(I);
8484 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8485 /// is to leave as a vector operation.
8486 static bool CheapToScalarize(Value *V, bool isConstant) {
8487 if (isa<ConstantAggregateZero>(V))
8489 if (ConstantPacked *C = dyn_cast<ConstantPacked>(V)) {
8490 if (isConstant) return true;
8491 // If all elts are the same, we can extract.
8492 Constant *Op0 = C->getOperand(0);
8493 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8494 if (C->getOperand(i) != Op0)
8498 Instruction *I = dyn_cast<Instruction>(V);
8499 if (!I) return false;
8501 // Insert element gets simplified to the inserted element or is deleted if
8502 // this is constant idx extract element and its a constant idx insertelt.
8503 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8504 isa<ConstantInt>(I->getOperand(2)))
8506 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8508 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8509 if (BO->hasOneUse() &&
8510 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8511 CheapToScalarize(BO->getOperand(1), isConstant)))
8513 if (CmpInst *CI = dyn_cast<CmpInst>(I))
8514 if (CI->hasOneUse() &&
8515 (CheapToScalarize(CI->getOperand(0), isConstant) ||
8516 CheapToScalarize(CI->getOperand(1), isConstant)))
8522 /// getShuffleMask - Read and decode a shufflevector mask. It turns undef
8523 /// elements into values that are larger than the #elts in the input.
8524 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8525 unsigned NElts = SVI->getType()->getNumElements();
8526 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8527 return std::vector<unsigned>(NElts, 0);
8528 if (isa<UndefValue>(SVI->getOperand(2)))
8529 return std::vector<unsigned>(NElts, 2*NElts);
8531 std::vector<unsigned> Result;
8532 const ConstantPacked *CP = cast<ConstantPacked>(SVI->getOperand(2));
8533 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8534 if (isa<UndefValue>(CP->getOperand(i)))
8535 Result.push_back(NElts*2); // undef -> 8
8537 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8541 /// FindScalarElement - Given a vector and an element number, see if the scalar
8542 /// value is already around as a register, for example if it were inserted then
8543 /// extracted from the vector.
8544 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8545 assert(isa<PackedType>(V->getType()) && "Not looking at a vector?");
8546 const PackedType *PTy = cast<PackedType>(V->getType());
8547 unsigned Width = PTy->getNumElements();
8548 if (EltNo >= Width) // Out of range access.
8549 return UndefValue::get(PTy->getElementType());
8551 if (isa<UndefValue>(V))
8552 return UndefValue::get(PTy->getElementType());
8553 else if (isa<ConstantAggregateZero>(V))
8554 return Constant::getNullValue(PTy->getElementType());
8555 else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V))
8556 return CP->getOperand(EltNo);
8557 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8558 // If this is an insert to a variable element, we don't know what it is.
8559 if (!isa<ConstantInt>(III->getOperand(2)))
8561 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8563 // If this is an insert to the element we are looking for, return the
8566 return III->getOperand(1);
8568 // Otherwise, the insertelement doesn't modify the value, recurse on its
8570 return FindScalarElement(III->getOperand(0), EltNo);
8571 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8572 unsigned InEl = getShuffleMask(SVI)[EltNo];
8574 return FindScalarElement(SVI->getOperand(0), InEl);
8575 else if (InEl < Width*2)
8576 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8578 return UndefValue::get(PTy->getElementType());
8581 // Otherwise, we don't know.
8585 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8587 // If packed val is undef, replace extract with scalar undef.
8588 if (isa<UndefValue>(EI.getOperand(0)))
8589 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8591 // If packed val is constant 0, replace extract with scalar 0.
8592 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8593 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8595 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
8596 // If packed val is constant with uniform operands, replace EI
8597 // with that operand
8598 Constant *op0 = C->getOperand(0);
8599 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8600 if (C->getOperand(i) != op0) {
8605 return ReplaceInstUsesWith(EI, op0);
8608 // If extracting a specified index from the vector, see if we can recursively
8609 // find a previously computed scalar that was inserted into the vector.
8610 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8611 // This instruction only demands the single element from the input vector.
8612 // If the input vector has a single use, simplify it based on this use
8614 uint64_t IndexVal = IdxC->getZExtValue();
8615 if (EI.getOperand(0)->hasOneUse()) {
8617 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8620 EI.setOperand(0, V);
8625 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8626 return ReplaceInstUsesWith(EI, Elt);
8629 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8630 if (I->hasOneUse()) {
8631 // Push extractelement into predecessor operation if legal and
8632 // profitable to do so
8633 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8634 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8635 if (CheapToScalarize(BO, isConstantElt)) {
8636 ExtractElementInst *newEI0 =
8637 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8638 EI.getName()+".lhs");
8639 ExtractElementInst *newEI1 =
8640 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8641 EI.getName()+".rhs");
8642 InsertNewInstBefore(newEI0, EI);
8643 InsertNewInstBefore(newEI1, EI);
8644 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8646 } else if (isa<LoadInst>(I)) {
8647 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8648 PointerType::get(EI.getType()), EI);
8649 GetElementPtrInst *GEP =
8650 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8651 InsertNewInstBefore(GEP, EI);
8652 return new LoadInst(GEP);
8655 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8656 // Extracting the inserted element?
8657 if (IE->getOperand(2) == EI.getOperand(1))
8658 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8659 // If the inserted and extracted elements are constants, they must not
8660 // be the same value, extract from the pre-inserted value instead.
8661 if (isa<Constant>(IE->getOperand(2)) &&
8662 isa<Constant>(EI.getOperand(1))) {
8663 AddUsesToWorkList(EI);
8664 EI.setOperand(0, IE->getOperand(0));
8667 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8668 // If this is extracting an element from a shufflevector, figure out where
8669 // it came from and extract from the appropriate input element instead.
8670 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8671 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8673 if (SrcIdx < SVI->getType()->getNumElements())
8674 Src = SVI->getOperand(0);
8675 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8676 SrcIdx -= SVI->getType()->getNumElements();
8677 Src = SVI->getOperand(1);
8679 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8681 return new ExtractElementInst(Src, SrcIdx);
8688 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8689 /// elements from either LHS or RHS, return the shuffle mask and true.
8690 /// Otherwise, return false.
8691 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8692 std::vector<Constant*> &Mask) {
8693 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8694 "Invalid CollectSingleShuffleElements");
8695 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8697 if (isa<UndefValue>(V)) {
8698 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
8700 } else if (V == LHS) {
8701 for (unsigned i = 0; i != NumElts; ++i)
8702 Mask.push_back(ConstantInt::get(Type::UIntTy, i));
8704 } else if (V == RHS) {
8705 for (unsigned i = 0; i != NumElts; ++i)
8706 Mask.push_back(ConstantInt::get(Type::UIntTy, i+NumElts));
8708 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8709 // If this is an insert of an extract from some other vector, include it.
8710 Value *VecOp = IEI->getOperand(0);
8711 Value *ScalarOp = IEI->getOperand(1);
8712 Value *IdxOp = IEI->getOperand(2);
8714 if (!isa<ConstantInt>(IdxOp))
8716 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8718 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8719 // Okay, we can handle this if the vector we are insertinting into is
8721 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8722 // If so, update the mask to reflect the inserted undef.
8723 Mask[InsertedIdx] = UndefValue::get(Type::UIntTy);
8726 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8727 if (isa<ConstantInt>(EI->getOperand(1)) &&
8728 EI->getOperand(0)->getType() == V->getType()) {
8729 unsigned ExtractedIdx =
8730 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8732 // This must be extracting from either LHS or RHS.
8733 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8734 // Okay, we can handle this if the vector we are insertinting into is
8736 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8737 // If so, update the mask to reflect the inserted value.
8738 if (EI->getOperand(0) == LHS) {
8739 Mask[InsertedIdx & (NumElts-1)] =
8740 ConstantInt::get(Type::UIntTy, ExtractedIdx);
8742 assert(EI->getOperand(0) == RHS);
8743 Mask[InsertedIdx & (NumElts-1)] =
8744 ConstantInt::get(Type::UIntTy, ExtractedIdx+NumElts);
8753 // TODO: Handle shufflevector here!
8758 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8759 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8760 /// that computes V and the LHS value of the shuffle.
8761 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8763 assert(isa<PackedType>(V->getType()) &&
8764 (RHS == 0 || V->getType() == RHS->getType()) &&
8765 "Invalid shuffle!");
8766 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8768 if (isa<UndefValue>(V)) {
8769 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
8771 } else if (isa<ConstantAggregateZero>(V)) {
8772 Mask.assign(NumElts, ConstantInt::get(Type::UIntTy, 0));
8774 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8775 // If this is an insert of an extract from some other vector, include it.
8776 Value *VecOp = IEI->getOperand(0);
8777 Value *ScalarOp = IEI->getOperand(1);
8778 Value *IdxOp = IEI->getOperand(2);
8780 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8781 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8782 EI->getOperand(0)->getType() == V->getType()) {
8783 unsigned ExtractedIdx =
8784 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8785 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8787 // Either the extracted from or inserted into vector must be RHSVec,
8788 // otherwise we'd end up with a shuffle of three inputs.
8789 if (EI->getOperand(0) == RHS || RHS == 0) {
8790 RHS = EI->getOperand(0);
8791 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8792 Mask[InsertedIdx & (NumElts-1)] =
8793 ConstantInt::get(Type::UIntTy, NumElts+ExtractedIdx);
8798 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8799 // Everything but the extracted element is replaced with the RHS.
8800 for (unsigned i = 0; i != NumElts; ++i) {
8801 if (i != InsertedIdx)
8802 Mask[i] = ConstantInt::get(Type::UIntTy, NumElts+i);
8807 // If this insertelement is a chain that comes from exactly these two
8808 // vectors, return the vector and the effective shuffle.
8809 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8810 return EI->getOperand(0);
8815 // TODO: Handle shufflevector here!
8817 // Otherwise, can't do anything fancy. Return an identity vector.
8818 for (unsigned i = 0; i != NumElts; ++i)
8819 Mask.push_back(ConstantInt::get(Type::UIntTy, i));
8823 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8824 Value *VecOp = IE.getOperand(0);
8825 Value *ScalarOp = IE.getOperand(1);
8826 Value *IdxOp = IE.getOperand(2);
8828 // If the inserted element was extracted from some other vector, and if the
8829 // indexes are constant, try to turn this into a shufflevector operation.
8830 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8831 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8832 EI->getOperand(0)->getType() == IE.getType()) {
8833 unsigned NumVectorElts = IE.getType()->getNumElements();
8834 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8835 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8837 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8838 return ReplaceInstUsesWith(IE, VecOp);
8840 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8841 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8843 // If we are extracting a value from a vector, then inserting it right
8844 // back into the same place, just use the input vector.
8845 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8846 return ReplaceInstUsesWith(IE, VecOp);
8848 // We could theoretically do this for ANY input. However, doing so could
8849 // turn chains of insertelement instructions into a chain of shufflevector
8850 // instructions, and right now we do not merge shufflevectors. As such,
8851 // only do this in a situation where it is clear that there is benefit.
8852 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8853 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8854 // the values of VecOp, except then one read from EIOp0.
8855 // Build a new shuffle mask.
8856 std::vector<Constant*> Mask;
8857 if (isa<UndefValue>(VecOp))
8858 Mask.assign(NumVectorElts, UndefValue::get(Type::UIntTy));
8860 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8861 Mask.assign(NumVectorElts, ConstantInt::get(Type::UIntTy,
8864 Mask[InsertedIdx] = ConstantInt::get(Type::UIntTy, ExtractedIdx);
8865 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8866 ConstantPacked::get(Mask));
8869 // If this insertelement isn't used by some other insertelement, turn it
8870 // (and any insertelements it points to), into one big shuffle.
8871 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8872 std::vector<Constant*> Mask;
8874 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8875 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8876 // We now have a shuffle of LHS, RHS, Mask.
8877 return new ShuffleVectorInst(LHS, RHS, ConstantPacked::get(Mask));
8886 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8887 Value *LHS = SVI.getOperand(0);
8888 Value *RHS = SVI.getOperand(1);
8889 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8891 bool MadeChange = false;
8893 // Undefined shuffle mask -> undefined value.
8894 if (isa<UndefValue>(SVI.getOperand(2)))
8895 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8897 // TODO: If we have shuffle(x, undef, mask) and any elements of mask refer to
8898 // the undef, change them to undefs.
8900 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8901 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8902 if (LHS == RHS || isa<UndefValue>(LHS)) {
8903 if (isa<UndefValue>(LHS) && LHS == RHS) {
8904 // shuffle(undef,undef,mask) -> undef.
8905 return ReplaceInstUsesWith(SVI, LHS);
8908 // Remap any references to RHS to use LHS.
8909 std::vector<Constant*> Elts;
8910 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8912 Elts.push_back(UndefValue::get(Type::UIntTy));
8914 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8915 (Mask[i] < e && isa<UndefValue>(LHS)))
8916 Mask[i] = 2*e; // Turn into undef.
8918 Mask[i] &= (e-1); // Force to LHS.
8919 Elts.push_back(ConstantInt::get(Type::UIntTy, Mask[i]));
8922 SVI.setOperand(0, SVI.getOperand(1));
8923 SVI.setOperand(1, UndefValue::get(RHS->getType()));
8924 SVI.setOperand(2, ConstantPacked::get(Elts));
8925 LHS = SVI.getOperand(0);
8926 RHS = SVI.getOperand(1);
8930 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
8931 bool isLHSID = true, isRHSID = true;
8933 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8934 if (Mask[i] >= e*2) continue; // Ignore undef values.
8935 // Is this an identity shuffle of the LHS value?
8936 isLHSID &= (Mask[i] == i);
8938 // Is this an identity shuffle of the RHS value?
8939 isRHSID &= (Mask[i]-e == i);
8942 // Eliminate identity shuffles.
8943 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
8944 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
8946 // If the LHS is a shufflevector itself, see if we can combine it with this
8947 // one without producing an unusual shuffle. Here we are really conservative:
8948 // we are absolutely afraid of producing a shuffle mask not in the input
8949 // program, because the code gen may not be smart enough to turn a merged
8950 // shuffle into two specific shuffles: it may produce worse code. As such,
8951 // we only merge two shuffles if the result is one of the two input shuffle
8952 // masks. In this case, merging the shuffles just removes one instruction,
8953 // which we know is safe. This is good for things like turning:
8954 // (splat(splat)) -> splat.
8955 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
8956 if (isa<UndefValue>(RHS)) {
8957 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
8959 std::vector<unsigned> NewMask;
8960 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
8962 NewMask.push_back(2*e);
8964 NewMask.push_back(LHSMask[Mask[i]]);
8966 // If the result mask is equal to the src shuffle or this shuffle mask, do
8968 if (NewMask == LHSMask || NewMask == Mask) {
8969 std::vector<Constant*> Elts;
8970 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
8971 if (NewMask[i] >= e*2) {
8972 Elts.push_back(UndefValue::get(Type::UIntTy));
8974 Elts.push_back(ConstantInt::get(Type::UIntTy, NewMask[i]));
8977 return new ShuffleVectorInst(LHSSVI->getOperand(0),
8978 LHSSVI->getOperand(1),
8979 ConstantPacked::get(Elts));
8984 return MadeChange ? &SVI : 0;
8989 void InstCombiner::removeFromWorkList(Instruction *I) {
8990 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
8995 /// TryToSinkInstruction - Try to move the specified instruction from its
8996 /// current block into the beginning of DestBlock, which can only happen if it's
8997 /// safe to move the instruction past all of the instructions between it and the
8998 /// end of its block.
8999 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9000 assert(I->hasOneUse() && "Invariants didn't hold!");
9002 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9003 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9005 // Do not sink alloca instructions out of the entry block.
9006 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
9009 // We can only sink load instructions if there is nothing between the load and
9010 // the end of block that could change the value.
9011 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9012 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9014 if (Scan->mayWriteToMemory())
9018 BasicBlock::iterator InsertPos = DestBlock->begin();
9019 while (isa<PHINode>(InsertPos)) ++InsertPos;
9021 I->moveBefore(InsertPos);
9026 /// OptimizeConstantExpr - Given a constant expression and target data layout
9027 /// information, symbolically evaluate the constant expr to something simpler
9029 static Constant *OptimizeConstantExpr(ConstantExpr *CE, const TargetData *TD) {
9032 Constant *Ptr = CE->getOperand(0);
9033 if (CE->getOpcode() == Instruction::GetElementPtr && Ptr->isNullValue() &&
9034 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
9035 // If this is a constant expr gep that is effectively computing an
9036 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
9037 bool isFoldableGEP = true;
9038 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
9039 if (!isa<ConstantInt>(CE->getOperand(i)))
9040 isFoldableGEP = false;
9041 if (isFoldableGEP) {
9042 std::vector<Value*> Ops(CE->op_begin()+1, CE->op_end());
9043 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(), Ops);
9044 Constant *C = ConstantInt::get(TD->getIntPtrType(), Offset);
9045 return ConstantExpr::getIntToPtr(C, CE->getType());
9053 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9054 /// all reachable code to the worklist.
9056 /// This has a couple of tricks to make the code faster and more powerful. In
9057 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9058 /// them to the worklist (this significantly speeds up instcombine on code where
9059 /// many instructions are dead or constant). Additionally, if we find a branch
9060 /// whose condition is a known constant, we only visit the reachable successors.
9062 static void AddReachableCodeToWorklist(BasicBlock *BB,
9063 std::set<BasicBlock*> &Visited,
9064 std::vector<Instruction*> &WorkList,
9065 const TargetData *TD) {
9066 // We have now visited this block! If we've already been here, bail out.
9067 if (!Visited.insert(BB).second) return;
9069 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9070 Instruction *Inst = BBI++;
9072 // DCE instruction if trivially dead.
9073 if (isInstructionTriviallyDead(Inst)) {
9075 DOUT << "IC: DCE: " << *Inst;
9076 Inst->eraseFromParent();
9080 // ConstantProp instruction if trivially constant.
9081 if (Constant *C = ConstantFoldInstruction(Inst)) {
9082 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
9083 C = OptimizeConstantExpr(CE, TD);
9084 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9085 Inst->replaceAllUsesWith(C);
9087 Inst->eraseFromParent();
9091 WorkList.push_back(Inst);
9094 // Recursively visit successors. If this is a branch or switch on a constant,
9095 // only visit the reachable successor.
9096 TerminatorInst *TI = BB->getTerminator();
9097 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9098 if (BI->isConditional() && isa<ConstantBool>(BI->getCondition())) {
9099 bool CondVal = cast<ConstantBool>(BI->getCondition())->getValue();
9100 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
9104 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9105 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9106 // See if this is an explicit destination.
9107 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9108 if (SI->getCaseValue(i) == Cond) {
9109 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
9113 // Otherwise it is the default destination.
9114 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
9119 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9120 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
9123 bool InstCombiner::runOnFunction(Function &F) {
9124 bool Changed = false;
9125 TD = &getAnalysis<TargetData>();
9128 // Do a depth-first traversal of the function, populate the worklist with
9129 // the reachable instructions. Ignore blocks that are not reachable. Keep
9130 // track of which blocks we visit.
9131 std::set<BasicBlock*> Visited;
9132 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
9134 // Do a quick scan over the function. If we find any blocks that are
9135 // unreachable, remove any instructions inside of them. This prevents
9136 // the instcombine code from having to deal with some bad special cases.
9137 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9138 if (!Visited.count(BB)) {
9139 Instruction *Term = BB->getTerminator();
9140 while (Term != BB->begin()) { // Remove instrs bottom-up
9141 BasicBlock::iterator I = Term; --I;
9143 DOUT << "IC: DCE: " << *I;
9146 if (!I->use_empty())
9147 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9148 I->eraseFromParent();
9153 while (!WorkList.empty()) {
9154 Instruction *I = WorkList.back(); // Get an instruction from the worklist
9155 WorkList.pop_back();
9157 // Check to see if we can DCE the instruction.
9158 if (isInstructionTriviallyDead(I)) {
9159 // Add operands to the worklist.
9160 if (I->getNumOperands() < 4)
9161 AddUsesToWorkList(*I);
9164 DOUT << "IC: DCE: " << *I;
9166 I->eraseFromParent();
9167 removeFromWorkList(I);
9171 // Instruction isn't dead, see if we can constant propagate it.
9172 if (Constant *C = ConstantFoldInstruction(I)) {
9173 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
9174 C = OptimizeConstantExpr(CE, TD);
9175 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9177 // Add operands to the worklist.
9178 AddUsesToWorkList(*I);
9179 ReplaceInstUsesWith(*I, C);
9182 I->eraseFromParent();
9183 removeFromWorkList(I);
9187 // See if we can trivially sink this instruction to a successor basic block.
9188 if (I->hasOneUse()) {
9189 BasicBlock *BB = I->getParent();
9190 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9191 if (UserParent != BB) {
9192 bool UserIsSuccessor = false;
9193 // See if the user is one of our successors.
9194 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9195 if (*SI == UserParent) {
9196 UserIsSuccessor = true;
9200 // If the user is one of our immediate successors, and if that successor
9201 // only has us as a predecessors (we'd have to split the critical edge
9202 // otherwise), we can keep going.
9203 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9204 next(pred_begin(UserParent)) == pred_end(UserParent))
9205 // Okay, the CFG is simple enough, try to sink this instruction.
9206 Changed |= TryToSinkInstruction(I, UserParent);
9210 // Now that we have an instruction, try combining it to simplify it...
9211 if (Instruction *Result = visit(*I)) {
9213 // Should we replace the old instruction with a new one?
9215 DOUT << "IC: Old = " << *I
9216 << " New = " << *Result;
9218 // Everything uses the new instruction now.
9219 I->replaceAllUsesWith(Result);
9221 // Push the new instruction and any users onto the worklist.
9222 WorkList.push_back(Result);
9223 AddUsersToWorkList(*Result);
9225 // Move the name to the new instruction first...
9226 std::string OldName = I->getName(); I->setName("");
9227 Result->setName(OldName);
9229 // Insert the new instruction into the basic block...
9230 BasicBlock *InstParent = I->getParent();
9231 BasicBlock::iterator InsertPos = I;
9233 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9234 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9237 InstParent->getInstList().insert(InsertPos, Result);
9239 // Make sure that we reprocess all operands now that we reduced their
9241 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9242 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9243 WorkList.push_back(OpI);
9245 // Instructions can end up on the worklist more than once. Make sure
9246 // we do not process an instruction that has been deleted.
9247 removeFromWorkList(I);
9249 // Erase the old instruction.
9250 InstParent->getInstList().erase(I);
9252 DOUT << "IC: MOD = " << *I;
9254 // If the instruction was modified, it's possible that it is now dead.
9255 // if so, remove it.
9256 if (isInstructionTriviallyDead(I)) {
9257 // Make sure we process all operands now that we are reducing their
9259 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
9260 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
9261 WorkList.push_back(OpI);
9263 // Instructions may end up in the worklist more than once. Erase all
9264 // occurrences of this instruction.
9265 removeFromWorkList(I);
9266 I->eraseFromParent();
9268 WorkList.push_back(Result);
9269 AddUsersToWorkList(*Result);
9279 FunctionPass *llvm::createInstructionCombiningPass() {
9280 return new InstCombiner();